THE BEHAVIORAL AND BRAIN SCIENCES (1981) 4, 609-650

Printed in the United States ot America

Precis of GallistePs The organization of action: A new synthesis

C. R. Gatlistel
Department of Psychology, University of Pennsylvania, Philadelphia, Pa. 19104

Abstract: The book describes three elementary units of action - the reflex, the oscillator, and the servomechanism - and the principles by which they are combined to make complex units. The combining of elementary units to make complex units gives behavior and the neural circuitry underlying behavior a hierarchical structure. Circuits at higher levels govern the operation of lower circuits by selective potentiation and depotentiation: by regulating the potential for operation in lower circuits - raising the potential for some and lowering it for others - a higher unit establishes the overall pattern to be exhibited in the combined operations of the lower units, while leaving it to the lower units to determine the details of the implementation of this pattern. A theory of motivation of the kind long championed by ethologists and physiological psychologists grows out of the notion of the selective potentiation and depotentiation of hierarchically structured units of behavior. The question then becomes how this conception of motivation may be integrated with a conception of a learned representation of the world to yield a theory of how animals produce novel motivated behavioral sequences based on learned representations. The representation of space, which is central to the behavior of organisms at least as lowly as the digger wasp, is a promising domain for the investigation of this question. Another promising area is the representation of skilled movements, such as handwriting. A number of issues in cognitive and developmental psychology are illumined by this theory about the organization of action. Among these are the degrees-of-freedom problem, the role of practice in development and in the learning of skilled action, and the role of mental rotation of spatial representations. Keywords: action; hierarchical structure; motivation; motor organization; oscillators; potentiation; reflexes; representations; servomechanisms; units of behavior

In his penetrating history of nineteenth-century neuropsychology, Robert M. Young (1970) observes that "the most fundamental and perplexing problem in psychology has been, and remains, the lack of an agreed set of units for analysis comparable to the elementary particles in physics and the periodic table of elements in chemistry" (p. viii). In The organization of action: A new synthesis I build the argument around a description of three distinct kinds of elementary units of behavior - the reflex, the oscillator, and the servomechanism. My purpose is threefold. I want to introduce students to these units by reprinting and explicating classic papers that delineate the properties of these units and the principles governing their interaction. I also want to place the traditional theory of motivation held by ethologists and many physiological psychologists (the drive or instinct theory) on a firm foundation, by showing that it emerges inescapably from a consideration of the principles by which complex behavior is built up from simple pieces of behavior. Lastly, I want to begin the process of integrating this view with modern work in cognitive and developmental psychology. A concern for discovering appropriate units of analysis runs through the whole book.
e 1981 Cambridge University Press 0 H0-525X/81/040609-42/$04.00/0

Metaphysics

I begin the book with a short exposition of the materialist metaphysics that is implicit in most modern behavioral neurobiology. This metaphysical postulate is that the central nervous system is the organ of thought and action, in just the sense in which the heart is the organ of circulation. The explanation of the patterns that we see in animal action, no matter which animal and no matter how purposive and intelligent the pattern, is to be sought in the anatomical structure and physiological functioning of the animal's central nervous system. While most but not all (cf. Eccles 1980) neurobehavioral scientists probably subscribe to this view today, it is seldom made explicit, perhaps. because modern philosophy has not found any very satisfying way to reconcile it with our conception of ourselves as morally responsible actors. I make the materialist assumption explicit in the hope that the resultant underlying metaphysical tension will add to the interest of the rest of the book. A commitment to a reductionist analysis - an explanation of behavior first in terms of anatomy and physiology and ultimately in terms of physics and chemistry - is latent in a materialist stance. Despite my
609

Gallistel: Organization of action reductionist commitment, the book contains relatively little anatomy and neurophysiology. A reductionist analysis is premature until the analysis of behavioral phenomena has identified suitable behavioral units and determined the principles to which the physiological processes underlying those units must adhere. You could not have a molecular explanation of the phenomena of inheritance in the absence of the concept of a gene; no Mendel: no Watson and Crick.
The reflex

The second chapter begins with the work of one of the greatest of behavioral physiologists, Sir Charles Sherrington. Sherrington studied behavior - reflexive behavior with a view to inferring the underlying neurophysiology; this, according to my definition, is behavioral physiology. In the first chapter of The integrative action of the nervous system, which is reprinted in edited form, Sherrington (1906, 1947) established several pivotal conceptions: He laid down the criteria for recognizing something as an elementary unit of behavior. A unit of behavior is a naturally occurring effector action (muscular or glandular) for which we can specify or hope to specify the effectors involved, the pathways by which the actioninitiating signals are conducted to the effectors, and the source of these signals. In short, a unit of behavior has three kinds of sub-behavioral constituents - effectors, conductors, and initiators. A unit is elementary if it cannot be analyzed into constituents that are themselves units of behavior. The tendon jerk reflex is a unit of behavior (though whether it normally functions as a reflex still remains to be decided). The initiator constituents are the stretch receptors in the intrafusal organs embedded in the muscle; the conductors are the la afferents from the muscle to the spinal cord and the a-motor neurons from the spinal cord back to the muscle; and the effectors are the muscle fibers. The tendon jerk reflex is an elementary unit of behavior because its constituents are not themselves units of behavior. The motor unit - an a-motor neuron and the muscle fibers it innervates - is not a unit of behavior, because this sub-behavioral unit of analysis lacks an initiator constituent. In mammals at least, a-motor neurons are not under natural circumstances the source of the action-initiating signals. Sensory receptors are the source of action-initiating signals in the reflex, but by themselves they are also not a unit of behavior, because they lack the effector constituent (and often any conductor constituent, as well). The coordinated stepping of the limbs in locomotion is also a unit of behavior, but not an elementary unit, because it may be analyzed into constituents - the stepping of each leg - that are units of behavior in their own right. Sherrington made the following inferences: (1) To account for the specificity of the stimuli required to elicit a given reflex, it is necessary to suppose that the receptors act as filters selectively tuned to certain kinds of stimuli. (2) To account for temporal summation seen in the elicitation of reflexes, it is necessary to suppose that there exist in the reflex arc physiological processes whose sluggish rise and fall make them capable of summating the effects of signals arriving several 610
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

seconds apart. (3) To account for the fact that two signals starting out from different receptors may summate to produce a signal to the effectors, it is necessary to suppose that the reflex arc has a region where the effects of signals arriving via different pathways may be combined. (4) From the fact that a signal originating at one site may cancel the excitatory effect of a signal originating at another site, it is necessary to suppose that the summation of incoming signals in the CNS (central nervous system) may be either additive or subtractive. (5) From the fact that two inputs, neither of which is alone capable of eliciting any effector action, may together elicit a vigorous action, it is necessary to suppose that these summation processes occur beneath a threshold, a signal going to the effector only when the sum exceeds the threshold. Sherrington made the further (and then controversial) assumption that the recently propounded cell theory applied to the nervous system as -well. Putting this assumption together with the above five inferences, he arrived at his conception of the synapse - a brilliantly successful example of inference from behavior to underlying physiology, which has stood the test of time very well indeed.
Combinatorial principles

Chapter 3 expounds Sherrington's combinatorial principles. The first of these was the principle of the common path. In propounding this principle, Sherrington called attention to the fact that higher units in the nervous system act by way of a common pool of lower units. At the level of elementary units, this means that all the reflexes for a given limb make use of the same motor neurons. But Sherrington pointed out (on anatomical grounds) that the same principle applied at higher levels. Higher levels did not have direct access to the motor neurons. They achieved their influence on effector action by way of the control they exerted over intermediate units. These intermediate units were also common paths, because they were controlled by diverse and competing higher-level signals. I term this the lattice hierarchy principle because when the control of lower units by higher units is diagrammed, the diagram looks like a ramshackle lattice (Figure 1). Sherrington goes on to identify various coordinative phenomena such as induction and irradiation and to

Figure 1. The distinction between a lattice hierarchy (A) and a partition hierarchy (B). The system underlying the generation of animal action is a lattice hierarchy.

Gallistel: Organization of action explain them in terms of nonlinear summation at synapses. He points out that the common path (lattice hierarchy) principle requires mechanisms for determining which of the many competing higher units has control of a common pool of subordinate units at any moment. He recognizes two principles mediating this intralevel competition for control: mutual facilitation between agonistic units and reciprocal inhibition between antagonistic units. He recognizes that these principles apply at the motor neuron level (reciprocal innervation of antagonist muscles) and at higher levels (reciprocal inhibition between competing reflexes). Lastly, Sherrington appeals to the reflex-chaining principle to explain sustained sequences of functionally coherent acts. I argue that this last principle, while certainly important in some instances, is not the magic wand it has been taken to be. It does not go very far toward explaining the structure of most complex behavioral sequences.
Oscillators

Sherrington thought of the coordinative function of the nervous system entirely in terms of the conduction of signals from receptors to effectors, with allowance made for the addition and subtraction of conducted signals in the CNS. To imagine that the CNS initiated signals in the absence of any sensory provocation no doubt smacked too much of vitalism. Although Sherrington was a dualist, he thought the soul initiated neural activity in the brain, not in the spinal cord. He knew that the rhythmic, patterned contraction of the heart muscle was driven by signals originating within the heart, but he did not consider (until much later) the notion that the other rhythmic effector actions, whose pervasiveness he clearly recognized, could be driven by signals from rhythmic signal generators (pacemakers) in the CNS. Sherrington's immediate successor in my pantheon of great behavioral physiologists is Erich von Hoist (1939), who did for the oscillator what Sherrington did for the reflex. He followed Graham Brown's (1914) lead in recognizing oscillators as a distinct kind of elementary unit. He also discovered the principle by which oscillator outputs combine and the principle by which oscillators interact. His paper "On the nature of order in the central nervous system" is reprinted and explicated in chapter 4. [See Selverston: "Are Central Pattern Generators Understandable?" BBS 3(4) 1980.] Von Hoist recognized that most rhythmic effector actions are not driven by rhythmic stimuli. The rhythmic waving of the fins in his favorite experimental preparation - a fish with a transection just above the medulla - is not driven by any stimulus that is specific to that action; it is spontaneous. The rhythmic signals to the effectors have their source in populations of pacemaker neurons in the CNS (or possibly in pacemaker circuits). [See Selverston: Are Central Pattern Generators Understandable?" BBS 3(4) 1980.] These neural metronomes may put out their neural ticks and tocks in the absence of any sensory input. An oscillator is an elementary unit of behavior in which a pacemaker is the source of signals that drive a rhythmic effector action. Von Hoist recognized that both simple (quasi-

sinusoidal) and complex rhythms were common. He showed by closely argued experimental observations that complex oscillations in fish fins resulted from the superimposition of the outputs of two or more simple oscillators. He went on to show the same phenomena in mammals, including man. Since the principle of Fourier synthesis - superimposing simple oscillations to achieve movements of arbitrary complexity - is unfamiliar to many neurobehavioral scientists, it receives a detailed explication in the section that introduces von Hoist's paper. Von Hoist observed, however, that not all perturbations in the rhythm of a fish fin could be accounted for by superimposition of oscillator outputs; some were the consequence of an interaction between oscillators. He realized that this interaction was what kept one oscillator coordinated with another. He studied the interaction (which he called the magnet effect) and adduced the principle of phase-dependent acceleration/deceleration. This principle specifies the manner in which an oscillator responds to a timing signal from another oscillator (either internal or external). The response to the timing signal depends on the phase of the receiving oscillator at the moment the signal is received. During some phases, the timing signal accelerates the receiving oscillator; during others it decelerates it. The plot of the receiving oscillator's acceleration/deceleration as a function of the phase at which the timing signal arrives has come to be called the phase-response curve. So far as I know, von Hoist was first to make such plots. He pointed out that the phase-response characteristics of an oscillator determine the phase relationship between sender and receiver. Von Hoist's inferences, like Sherrington's, have been amply confirmed by subsequent electrophysiological work. The concept of coupled oscillators - oscillators that maintain a phase relationship by the exchange of timing signals - is a cornerstone of modern explanations of the neurobiological basis for a variety of complex rhythmic behaviors. The realization of how pervasive rhythmic action is in behavior, a point that Sherrington himself stressed, grows every year.
Locomotion and the degrees-of-freedom problem

Chapter 5 maintains the theme taken up in chapter 4 by reprinting and explicating Wilson's (1966) paper on insect locomotion, which owed much to von Hoist's work. This gets a little ahead of the argument, because it introduces a complex unit of behavior, locomotion, before an account of the third kind of elementary unit, the servomechanism; but the connection between von Hoist's work and Wilson's is so direct that it seemed a shame to let the thread drop. Wilson's model, which is the direct ancestor of more recent models (e.g., Pearson 1976), is a coupled-oscillator model. Each of the insect's six legs has its own oscillator, which drives the stepping of the leg. In the version of the model that I elaborate, the leg-stepping commands from the oscillator are modified by two other elementary units of behavior. One is an inhibitory reflex, which prevents the initiation of a leg swing when other legs have failed to take up the load. The
THE BEHAVIORAL AND BRAIN SCIENCES (1981). 4

611

Gallistel: Organization of action other is a servomechanism, which adjusts the strength of the signals sent to the muscles of support and propulsion so as to compensate for variations in the load to be supported and moved. While there is some experimental justification for positing the existence of each of these components, their precise nature is rather unclear at present. I include them largely for didactic reasons - in order to illustrate the multilayered character of the hierarchical structure underlying complex units of behavior: the combination of the oscillator, the trigger-inhibiting reflex, and the load-compensating servomechanism forms a complex unit that controls the stepping of a single leg. The combination of six of these complex units, one for each leg, forms a more complex, higher-level unit that generates locomotion in an insect. Modern explanations of locomotion bear directly on a central problem in the understanding of action - the extraordinarily diverse ways in which seemingly unitary acts achieve expression in muscular contraction. A cockroach's walking straight ahead may seem a simple and unitary act, but the close observer of cockroaches will note that they have a great many different ways of doing it. There are so many different leg-stepping patterns that it is only a slight exaggeration to say that the cockroach never walks in exactly the same way twice in its life. By "exactly the same way" I mean an exact duplication of the timing and magnitude of contraction and relaxation in every one of the many muscles involved, over several complete stepping cycles. Animals, unlike most machines, do the same thing in many different ways, each way peculiarly suited to the circumstances of the moment. This is part of the reason we are loath to believe that animals really are machines. The problem is to account for this circumstance-adapted diversity of output without positing as many different neural curcuits as there are outputs to be explained. This is the degrees-of-freedom problem, or one aspect of it. Another aspect of the problem is to explain how higher levels may select among an endlessly rich set of possible lower-level outputs without getting bogged down in specifying every detail. A third aspect of the problem, one that has often preoccupied philosophers, is the question of why we are justified in regarding an act like walking as basically the same act from one occasion to the next when, upon more minute observation, there are so many differences from occasion to occasion. Modern models of locomotion shed light on all these aspects of the degrees-of-freedom problem. Wilson points out that four motor constancies emerge from an analysis of the bewilderingly diverse insect gaits: (1) If you focus your attention on just the legs on one side of the body, you find that in all gaits there is a wave of forward swings that progresses from back to front. First the hindleg swings, then the middle leg, then the foreleg. (2) The duration of a swing is almost constant; only the duration of the support phase of a step varies. (3) The interval between the swing of the hindleg and that of the middle leg is constant, as is the interval between the swing of the middle leg and the foreleg. (4) Turning now to the phase relation between legs, one finds a final constancy: legs on
612
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

C O

(-cIb~| -Time-

Figure 2. Black bars indicate the sequence of leg swings on the left (Li - left front, L2, L3) and right (R,, R2, R3) sides in slow-moving and scurrying insect gaits. Both gaits conform to four motor constancies: (1) the sequence of leg swings on a given side progresses from back to front (arrow a); (2) the duration of a leg swing is fixed (interval b)\ (3) the lags between hind- and middle-leg swings and between middleand front-leg swings arefixed(interval c); (4) opposing legs swing 180 out of phase (the ratio d/e - 1/2). The specification of the period from one swing of a leg to the next (interval e) exhausts the degrees of freedom in the parameterization of these gaits, given the constraints imposed by the four constancies. All other differences follow from this specification. The stepping frequency is 1/e (after Wilson 1966). opposite sides of the same body segment always step in alternation - that is, 180 out of phase. These four constancies leave only one degree of freedom - the stepping frequency, which is the reciprocal of the period from one swing of a leg to the next swing of the same leg (see Figure 2). Wilson shows that all the different basic gaits are the consequence of different values for this one variable. Modern models of insect locomotion translate each of the constancies into a structural feature of the model. The swing signal initiated by the pacemaker is assumed to have a fixed duration, hence the constancy of the swing interval. The pacemaker for the hindleg is assumed to send a timing signal to the pacemaker for the middle leg in such a way that the middle leg pacemaker lags behind the hindleg pacemaker by a fixed interval. Likewise for the foreleg pacemaker; it is made to lag behind the middle leg pacemaker by a fixed interval. The fixed-lag coupling between each oscillator and the ipsilateral oscillator to the rear of it explains both the back-to-front wave of swings and the fixed interswing intervals within the wave. Lastly, the pacemakers on opposite sides of the same body segment are assumed to be coupled 180 out of phase. The higher levels of the insect nervous system can specify any gait that circumstances require by means of a single command signal, a signal that sets the stepping frequency of all the oscillators to about the same value. This signal raises or lowers the weights on the metronomes, so to speak, thereby causing them to cycle slowly or rapidly. All the differences that may be noted among the basic gaits are a consequence of changing this one parameter in the locomotory machinery. This is why the small brain of an insect does not get bogged down in instructing the lower levels on how to go about walking at the rate it wants.

Gallistel: Organization of action This is also why we are justified in regarding the many different gaits as diverse manifestations of a single underlying system. Finally, the trigger-inhibiting reflex and the loadcompensating servomechanism act to adapt the basic gaits to the vicissitudes of walking, the slips and trips, the up-walls and down-walls. That is why the diversity of stepping patterns takes on the character of intelligence - the adaptation to momentary circumstances so as to preserve overall function. The best-known example of this adaptation is the insect's adaptation to amputation of its middle legs. The cockroach normally runs with a tripodal gait (bottom of Figure 2), in which the hind and front legs on one side are swung simultaneously. When its middle legs are amputated, rendering this pattern of hind and foreleg activation inadvisable, the roach switches immediately to a gait in which the hind and front legs on a side no longer swing simultaneously, the so-called diagonal gate of a tetrapod. This is wired-in intelligence, not learning.
Servomechanisms

Chapter 6 returns to the elementary units of behavior to describe a third kind, the servomechanism. A servomechanism, like a reflex, has sensory receptors as at least one source of action-initiating signals. What sets a servomechanism apart from a reflex is the fundamental role of negative feedback in the functioning of the servomechanism. The test of whether a unit is a servomechanism is to disrupt whatever feedback there may be and see if this disruption alters the operation of the unit. The characteristics of the neural circuitry underlying a servomechanism are comprehensible only if one bears in mind the role of the negative feedback from output to input. The clearest illustration of the difference between reflex and servomechanistic neural circuitry comes from work that I became aware of too late to put in the book. Baarsma & Collewijn (1974) subjected both the vestibulo-ocular reflex and the optokinetic reaction to analysis by linear systems methods. The vestibulo-ocular reflex is a counter-rotation of the eyes in response to a rotation of the head sensed by receptors in the semicircular canals. It functions to stabilize the retinal image during head rotations. There is no feedback from output to input, because eye rotation has no effect on the semicircular canals. The optokinetic reaction is a rotation of the eyes in the direction of image slippage. It is initiated by circuits in the retina sensitive to the slippage of the visual image. It, too, functions to stabilize the retinal image during head rotation. It, however, is a servomechanism. The rotation of the eyes in the direction of image slippage reduces the slippage that is the stimulus for the rotation. This negative feedback through the environment from the unit's output to its input is essential to the normal operation of the unit. Baarsma and Collewijn, working with the rabbit, measured the gain in both the vestibulo-ocular circuit and the optokinetic circuit; that is, they measured the angular velocity of the eye rotation elicited by a given angular velocity of either head rotation or image slippage. In order to compare directly the characteristics

of the neural circuits per se, it was necessary to prevent the negative feedback from eye rotation to image slippage, which is a normal part of the servocircuit's action. This they did by fixing one eye, which looked at a rotating drum, while leaving the other unfixed but covered, so that it was free to move, but saw nothing. This arrangement, where one eye sees while the other eye moves, forces the neural circuitry underlying the optokinetic reaction to function as a reflex, with no negative feedback from its output to its input. Baarsma and Collewijn found that the gain of the vestibulo-ocular reflex circuit was about .8, whereas the gain of the optokinetic servocircuit was 20-100. Because the output of the servocircuit normally reduces the stimulus for its own action, it is essential that a little stimulus cause a lot of action. This is what a gain of 20-100 means. When the servocircuit is allowed to function under the negative feedback conditions for which it was designed, the high gain in the receptor-to-effector path makes the unit's normal operating characteristic similar to the operating characteristic of the reflex. Both units reduce image slippage by a factor of between 5 and 10. Baarsma and Collewijn also found lower acceleration in the servocircuit than in the reflex circuit. Sluggish acceleration (poor highfrequency response) is necessary to preserve the servomechanism's stability - that is, to prevent its going into nonfunctional oscillations in response to sudden movements of the visual image. In short, neural circuitry subserving servomechanisms can be expected to have characteristics that distinguish it from circuitry subserving reflexes, which is why these should be recognized as distinct kinds of units of behavior. The taxic orienting behaviors in lower organisms were among the first servomechanisms studied by behavioral physiologists. One of the pleasures of writing my book was translating for chapter 6 a paper by Fraenkel (1927). He shows that the coastal snail is normally both negatively phototaxic (orients away from light) and negatively geotaxic (orients away from the pull of gravity). When the snail crawls on a vertical surface illuminated from one side, the outputs of these orienting servomechanisms combine additively, so that the snail's trail is the vector sum of a sideways-oriented (negatively phototaxic) crawl and a vertically oriented (negatively geotaxic) crawl. When the snail is both upside-down and underwater, something fascinating happens. The snail is no longer negatively phototaxic; it is positively phototaxic. Perhaps the most elegant demonstration of both the reversal of the sign of the phototaxis and the vector summation of geotaxic and phototaxic orientations comes from placing the snail on the inside bottom of a submerged glass cylinder illuminated obliquely (Figure 3). The snail crawls away from the light on an oblique angle to the cylinder's axis. As it mounts the side of the cylinder, the obliquity increases because the geotaxic orientation gets stronger. As the snail passes the point halfway up the side of the cylinder, it becomes upside-down, and the sign of the phototaxis reverses, so that it now crawls along the inside top of the cylinder back toward the light. Fraenkel argues that the snail's taxic orienting mechanisms go a long way toward explaining how it gets
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

613

Gallistel: Organization of action

obliquely incident light

Figure 3. Bohn's (1905) experiment demonstrating the vector summation of phototaxic and geotaxic orientations and the reversal of the sign of the phototaxic orientation when the snail is upside-down (and underwater).

from the sea floor to where it is usually found - in crevasses on coastal cliffs, a few meters above high tide. The reversal of the sign of the phototaxis moves the snail out along the ceiling of the underwater crevasses in the cliff. Since the reversal occurs only when the snail is both upside-down and underwater, the snail's progression comes to an end in the dark inner recess of the first crevasse it comes to above the reach of the sea. The reversal of the snail's phototaxis is the first of many examples of the principle of selective potentiation and depotentiation in the organization of behavior. The snail has neural circuitry for both positive and negative phototaxic orienting, but only one of these two opposing circuits is allowed to be active at any one time. When the snail is upright, the negatively phototaxic circuit has the potential to be active. Whether it is in fact active depends on whether there is a light gradient. When, however, the snail is upside-down and underwater, the negative phototaxis is depotentiated while the positive phototaxis is potentiated. By controlling the potential for activation in these orienting circuits, the higher levels of the nervous system forge a functionally cohesive sequence of orientations.
Reafference and efference copy

When an animal acts it generates sensory inputs that are a consequence of its own action. These are termed reafferences, to distinguish them from exafferences, which are sensory inputs that come from events in the outside world. These reafferences may activate a unit of behavior whose action interferes with the action of the original unit. When a fly turns, for example, it necessarily causes the visual pattern to slip across its faceted eye. This should activate the fly's optokinetic reaction, producing a turn that counteracts the original turn. Why does the fly's optokinetic reaction not in effect paralyze the fly by constantly opposing turns initiated by other units? Von Hoist & Mittelstaedt (1950), whose well-known paper "The reafference principle" is reprinted in chapter 7, show that interference from the optokinetic unit can be prevented by cancelling the reafferent input to that unit with a copy of the efferent (outgoing signal) that initiates a turn. Because the reafference is a consequence of the fly's own action, it may be predicted by the signal that commands the turn. The strength and sign (left-right) of the signal that commands a turn predict the strength and sign of the
614
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

signal that the turn will generate in the optical circuits that respond to image motion. A left-turn command will lead to a visual motion signal of opposite sign because a left turn causes the visual pattern to slip toward the fly's right. Therefore, adding a suitably scaled copy of the left-turn command to the afferent signal coming into the optokinetic circuit will cancel out the reafferent component of that input, leaving only the component that reflects movements of the world rather than movements of the fly. Von Hoist and Mittelstaedt go on to explore other kinds of interactions between efferent and reafferent signals, greatly expanding our sense of the role that servomechanistic units play in behavior. In the later parts of the paper, they focus on the perceptual consequences of the interplay between reafference and efference copies. The best-known demonstrations that the principle of cancellation by efference copy is at work in the human nervous system come from cases in which the response of the eye muscles to efferent signals is impaired by disease or drugs (e.g., curare). In such cases, the motion produced by the efference is less than it should be, hence also the reafference from the motion. When the efference copy is added to the weaker-than-usual reafferent, it overmatches it. Instead of cancelling out the reafferent, the efference copy inserts a signal of opposite sign into the sensory perceptual pathway. Hence, when individuals with a peripheral debilitation of the oculomotor apparatus command their eyes to look to the right, they perceive the world as jumping to the right. The efference copy has generated a perceptual effect all by itself, a perception for which there is no sensory source. Efference copies are low-level, neurophysiologically explicable instances of expectations. They are centrally generated signals that anticipate forthcoming sensory signals. They are, of course, unconscious expectations. But when these unconscious expectations are violated, the difference between expectation and outcome shows up in conscious experience. Von Hoist and Mittelstaedt give several other examples of this same phenomenon. [See also Gyr et al.: "Motor-Sensory Feedback and Geometry of Visual Space" BBS 2(1) 1979.]
Hierarchical structure

Chapters 2-7 consider different kinds of elementary units of behavior, the principles by which their outputs are combined, and the principles by which the units interact. Chapter 8 turns to a consideration of the hierarchical structure of the system. It reprints an edited version of a lengthy paper by Paul Weiss (1941) on the development of coordinated action in salamanders. Much of the developmental part has been edited out, because I wanted to emphasize what Weiss had to say about coordination itself. He stresses the hierarchical organization of the mechanisms of coordination, distinguishing six levels of coordination up to the level of simple acts: Level 1 is the motor unit - that is, the individual motor neuron and the muscle fibers it innervates. As already pointed out, this is a sub-behavioral unit of action. Level 2 is the coordination of the motor units that comprise a muscle. The size principle (Stein 1974), which asserts that smaller motor neurons

Gallistel: Organization of action commanding weaker contractile effects are always recruited first, is an example of a principle of coordination at this level. This is still a sub-behavioral level of analysis. Level 3 is the level of muscle groups - for example, the antagonistic muscles on either side of a hinge joint. Elementary units of behavior command muscle groups, so this is the first behavioral level of analysis. Level 4 is the level of the organ. The combination of an oscillator, a trigger-inhibiting reflex, and a load-compensating servomechanism, which together control the stepping of a single leg, is an example of a complex unit of behavior at this level. Level 5 is the level of the organ system. The unit that coordinates the stepping of all six legs in order to produce forward locomotion is an example of a unit at this level. Level 6 is the level of the organism as a whole. The units that produce oriented progressions, in which an orienting unit operates together with a locomotory unit, are examples of units at this level. Weiss's sixth level, the level of simple acts of the whole organism, may, for convenience at least, be regarded as the level at which the study of sensorimotor coordination leaves off and the study of motivated behavior begins. Motivated behavior is behavior in which many different acts are deployed in such a way as to achieve a goal or function that no one act achieves by itself. Oriented progressions are salient among the acts that are deployed in the appetitive phase of motivated behavior. Other acts - for example, the ingestion and rejection acts so carefully analyzed by Grill and Norgren (1978) in the rat - come into play during the consummatory phase. The centerpiece of Weiss s paper, for my purposes, is his compelling demonstration that locomotory systems at Level 5 of the hierarchy really are treated as units by higher levels of the salamander's nervous system. Weiss interchanged the forelimbs in larval salamanders, so that the animals' forelimbs were oriented backwards. These surgically translocated forelimbs were reinnervated, with the motor neurons appropriate to each kind of muscle finding their way to that muscle in the new backward-pointing limb. The result was that the pattern of motor neuron signals that would ordinarily step the forelimbs so as to propel the salamander forward caused, in the turned-around limbs, a pattern that was unaltered so far as the limbs themselves were concerned, but propelled the salamander backward, not forward. The normal salamander can walk both forward and backward. For these altered salamanders to do the same, they would have to instruct their hindlimbs to walk forward while instructing their forelimbs to walk backward, and vice versa. This they could not do, and never learned to do. The forelimbs and hindlimbs of Weiss's salamanders always worked at cross purposes. This is the most compelling demonstration I know of the unitary character of higher-level organization in the motor system. The salamander has a unit that generates forward locomotion; it also has a unit that generates backward locomotion. Each of these units generates a great diversity of stepping patterns. The salamander may call one, or it may call the other, but it cannot call half of one and half of the other. The diversity in its stepping patterns is delimited by the unitary hierarchical structure of the systems that generate the patterns. Weiss makes a severe indictment of most modern work on learning. Given that one is dealing with a hierarchically structured system for generating behavior, one of the most basic questions one can ask about the phenomena of plasticity is, at what level of the hierarchy does the alteration in structure occur? Until this question is answered, there is no hope of predicting the range of behavioral alterations consequent upon the learned alteration in structure. Wickens (1938, 1939), in his studies of response generalization in conditioned finger-withdrawal experiments, raised this question and demonstrated its pertinence. But the question has in general been left unanswered - indeed, unaddressed.
Manifestations of hierarchical structure

Chapter 9 surveys diverse material in order to show, first, the many ways in which the hierarchical structure of the system that generates action becomes manifest in behavioral experiments, and second, the many cases in which one may see the principle of selective potentiation and depotentiation at work. Transaction studies. On anatomical, phylogenetic, and ontogenetic (embryological) grounds, it is possible to distinguish five major divisions of the vertebrate central nervous system - the spinal cord, the hindbrain, the midbrain, the diencephalon, and the forebrain. Transection experiments study the behavioral capacities of animals whose nervous system has been cut at or near one of the boundaries between major divisions. These experiments reveal a rough correspondence between levels of organization in behavior and the major divisions of the CNS. In the spinal animal, units up to Weiss's Level 4 are intact. Fragments of Level 5 units are seen, but few units at this level are intact. In the hindbrain animal, many Level 5 units are intact or nearly so, but one does not see acts of the whole organism. In the midbrain animal, Level 6 units are intact. The animal walks, climbs, grooms, ingests, and so on. But one does not see motivated behavior. The acts are not deployed in such a way as to serve the animal's fundamental requirements - the preservation of the milieu interne, defense, and reproduction. In the animal whose diencephalon and immediately adjacent tissue have been spared by the transection or ablation, one sees motivated behavior. The devastating loss of the analytic and planning functions in the forebrain makes this motivated behavior very awkward and frequently poorly oriented, but the animal does deploy its limited means so as best to serve its nutritive, defensive, and reproductive goals. Brain stimulation. The lattice hierarchy conception was central to some of the most penetrating work ever done on the behavioral consequences of focal stimulation of the brainstem, the work by von Hoist and Ursula von St. Paul (1963). They used the lattice hierarchy conception to explain the multiple, seemingly diverse behavioral consequences of stimulating one and the same site, and to explain the fact that the same effects may
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

615

Gallistel: Organization of action be obtained from many different sites. They stress the supreme importance of a "level-adequate terminology" - that is, a description of the behavioral effect that accurately reflects the level of organization at which the units activated by the stimulation function. They point out that nothing but confusion and perplexity can result from lumping together in one functional category all the sites at which stimulation elicits, for example, some salient aspect of attack behavior. Some sites may function to coordinate that aspect and only that aspect of aggressive or defensive behavior, while other sites may govern a much wider domain, of which that piece of behavior is but a component. One must survey the behavioral effects in a variety of circumstances and know the animal's patterns of behavior thoroughly if one is to have any hope of making an adequate assessment of the behavioral function of the tissue excited by the stimulation. It cannot be said that most workers in brain stimulation have taken these strictures to heart. Flynn's work (e.g., 1972) on stimulation-elicited predation in cats provides superb demonstrations of selective potentiation in a motivational context. MacDonnell & Flynn (1966) showed that stimulating the cats diencephalon unilaterally at sites that induced complete predatory sequences of acts selectively potentiated several reflex components of the consummatory act on the contralateral side of the cat's face. In effect, the side of the cat contralateral to the site of diencephalic stimulation reacted to prey stimuli in a predatory fashion while the ipsilateral side reacted as it would in a cat not bent on capturing prey. This result argues strongly, I think, that diencephalic neural tissue achieves its motivating function by way of the selective potentiation of acts that subserve a common goal. Beagley & Holley (1977) have extended this finding to a learned act. They showed contralateral potentiation of a learned food-seeking response to a discriminative cue in the rat. The behavior of hybrids. Dilger's (1962) studies on nest building in hybrid lovebirds show the importance of inherited units of behavior in determining the outcome of learning experiments. He interbred two species of nest-building lovebirds, one of which carries material to the nest by tucking it in its feathers, the other carrying it in its bill. The hybrid offspring much preferred the feather-tucking approach, even though it never worked for them (since they did not execute the maneuver well and the material always fell out en route to the nest). Operant conditioning. Skinnerians have discovered that the inherited organization of behavior really does matter in operant conditioning. The Brelands (1961) were among the first to emphasize that operant behavior was synthesized from the preexisting speciesspecific units of behavior that tend to occur in connection with a given kind of reinforcer. When these units are not suited to the result the trainer wants to achieve, it is hard going, as Hineline and Rachlin (1969) found when they tried to condition a pigeon to peck a key to escape footshock. Not surprisingly, the birds tried to solve the problem with a unit of behavior ordinarily more suited to the goal of escaping painful stimuli to
616
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

the feet - flying. The work of Hearst & Jenkins (1975), Thorpe (1963), and Sevenster (1973) has further emphasized the role that preexisting higher level units of response organization play in instrumental learning. Development and recovery. The existence of higherlevel control over lower-level units of behavior is documented in Hall, Cramer, & Blass's (1977) studies on the development of suckling in rat pups, in Twitchell's (1970) studies of the development of grasping in human infants, and in Teitelbaum's many studies on the recovery of behavioral function following damage to the CNS in adults (cf. Teitelbaum 1971).
Motivation

Behaviorists, many neurobiologists, and some physiological psychologists mistrust the term "motivation." They seem to feel that when we really understand behavior we won't need to talk about motivation any more. I think they could not be more mistaken. The concept of motivation will play a central role in any satisfactory account of behavioral causation. There is no getting around the fact that animals pursue different goals on different occasions and that the goal an animal is pursuing determines how it will react, and even whether it will react at all, to a given stimulus. The cat in pursuit of prey reacts to a touch on the cheek by a turn toward the touch, and to a touch on the lips by a snapping open of the jaws. A cat not interested in prey turns away from a touch on the cheek and purses its lips to a touch on the lips. The study of motivation is the study of the goals animals pursue, the acts they can deploy in pursuit of those goals, and the processes that determine which goals will be pursued and how the acts will be deployed. I argue in chapter 10 that the principle of selective potentiation is central to understanding how motivating signals, arising in and around the diencephalon, impose goal-directedness on the animal's behavior. I follow the ethologists (Tinbergen 1951; Lorenz 1937) and many physiological psychologists (Lashley 1938; Morgan 1943; Stellar 1960) in arguing that central motive states (drives) increase the potential for certain acts and decrease the potential for others. They selectively favor the acts that subserve a common purpose. In this way, they establish the general tendency (the goal) of an animal's behavior, while leaving it to the activating effects of ambient conditions on lower-level units to determine just how the goal is pursued at any given moment. The behavior controlled by the higher, motivational levels of the action hierarchy is flexible, adaptable, purposive, and intelligent precisely because of the hierarchical arrangement. The higher levels lay down a frame or general direction for behavior by selectively potentiating coherent sets of behavioral options. The flexibility of the behavior, its variability from occasion to occasion, and its suitability to momentary and unforseeable circumstances reflect the fact that what higher levels establish are options, not requirements. The higher levels do not directly command muscles. They do not determine which muscles shall contract and when. They determine the set of general patterns of

Gallistel: Organization of action integrated sequences of simple acts that may be observed when a particular motivational state is in force. The imposition of a general plan by means of selective potentiation gives behavior its purposive quality, its long-term functional coherence. Each lower level in the hierarchy specifies more narrowly which of the many variations, subvariations, and sub-subvariations of the general plan will actually appear at a particular moment. Lower levels fill in or determine details within the general pattern. Since the stimuli and conditions that give rise to a particular drive are a familiar province of physiological psychology, I devote little space to this topic. I note that the kinds of processes that determine the waxing and waning of drive signals are analogous to the kinds of processes that give elementary units of behavior their distinctive characters. A drive may be elicited by a stimulus; it may have an endogenous oscillatory component; and, of course, the homeostatic drives have much to do with negative feedback regulation. Also, as Lorenz (1950) has emphasized, the potential for activating some acts may grow simply as a function of the interval since the act was last performed, quite independent of all other considerations, to the point where the act may occur in a behavioral vacuum - a situation in which there is no stimulus for the act and in which it serves no goal.
The oligarchical problem. Most propounders of the

thesis that the system generating behavior has a hierarchical structure have tried to specify units that constitute an oligarchy at the top - the major drives or major instincts (cf. Tinbergen 1951). These lists have proved an embarrassment. Many of the acts seen in seminatural conditions do not subserve any major drive (cf. Kavanau 1969; Leyhausen 1965); the conditions of their occurrence make such an assumption gratuitous. Also, as Lorenz (1950) first argued, the superordinatesubordinate relationship between units is frequently reversible. Rats deprived of water will run in order to gain access to water; but rats deprived of running will drink in order to gain access to a running wheel (Premack 1962). One must recognize that at motivational levels of function in higher vertebrates the hierarchy is labile. Which units subserve which is contextdependent. There is no oligarchy. There is a heterarchy.
Knowledge and action

What Young (1970) terms the sensorimotor approach to neuropsychology tries to solve the problem of the role of knowledge in the control of action by assuming that knowledge is an acquired alteration in sensorimotor connections. I argue in chapter 11 that this approach has been a failure. Knowledge often enters into the control of action in such a way as to produce behavior that the animal has never produced before and will never produce again. The female digger wasp, after digging an egg-laying burrow, gains knowledge of its location relative to surrounding landmarks during a survey flight that she makes before departing on her hunting expeditions. This knowledge enables her to find her way back to the burrow from an arbitrary

location, even one she is carried to inside a closed box (Thorpe 1950). To speak of her homing behavior as a habit or as an alteration in her response probabilities is a wave of the hand, because she orients correctly from release sites where she has never been. There is compelling evidence that she orients by reference to landmarks, not by reference to any stimulus beacon emanating from the burrow. The conclusion is inescapable that she has made a representation of the surrounding territory, a representation that she can use to orient toward the burrow from an arbitrarily chosen point within that territory. This conclusion poses two challenging questions: (1) How shall we describe her representation? (2) What is the process by which this representation enters into the control of her actions? [See also Griffin: "Prospects for a Cognitive Ethology" BBS 1(4) 1978.] To illustrate one approach to these questions, I review Deutsch's (1960) theory of map-referenced navigation in animals. Deutsch's theory of the representation assumes only two primitives - points (representations of distinctive landmarks) and connections between points (representations, roughly speaking, of the order in which the distinctive landmarks are encountered). Formally speaking, Deutsch assumes a network or graph-theoretic representation. Deutsch assumes that motivational signals enter the network via the representations of a goal object and propagate decrementally to all other points via the connections between points. The decremental propagation of the signal outwards from the goal point establishes a motivational gradient on the representation. At any one place in the territory, the animal will perceive only a subset of the points represented on its map. The motivational gradient establishes a rank ordering of these points on the basis of their motivational salience. That point whose representation on the map has the fewest connections separating it from the goal has the highest motivational salience. The animal orients toward the point of greatest salience and locomotes. This brings still more salient points into view. In this way the animal progresses to the goal. The above model is likely to make many behavioral neurobiologists squirm. It sounds so nonphysiological, so mentalistic. I review it because, for one thing, Deutsch (1960) actually built a working model. I do not know of any such attempt by theorists working in the sensorimotor tradition, and I believe that any such attempt would reveal the bankruptcy of their approach. It may also help to remember that the concept of the gene - self-replicating particle for the conveyance of structure-specifying information from one generation to the next - did not sound very chemical during the first half of this century. In any event, I argue in chapter 12 that Deutsch's model is probably inadequate, because animals' spatial representations are probably richer than he assumed. Deutsch's model assumes point-to-point navigation. The evidence for dead-reckoning navigation (cf. Carr & Watson 1908; Maier 1929; O'Keefe & Nadel 1978; Olton & Samuelson 1976; Tinkelpaugh 1932) implies a representation with a richer set of primitives, the primitives of Euclidean geometry. In dead-reckoning navigation, the animal does not perceive the point
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

617

Gallistel: Organization of action to which it is oriented; the orientation and the distance to be covered are derived from knowledge of the relation between the unperceived point toward which the course is set and other perceived points that do not lie on the course. The ability to derive the angle and distance of the required course would seem to require a representation that preserves angle, distance, collinearity, parallelism, and the other fundamental properties of Euclidean geometry. [See also Olton et al.: "Hippocampus, Space and Memory" BBS 2(3) 1979; and BBS multiple book review of O'Keefe & Nadel: "The Hippocampus as a Cognitive Map" BBS 2(4) 1979.] My original purpose was only to show that neurobiologically plausible conceptions of schemata could be imagined. The general outlines of the process for translating oscillatory letter schemata into movements are fairly clear. The point on the body that is to execute the trajectory must be set into at least two concurrent orthogonal oscillatory motions and a left-right linear motion. All the different letters are generated by modulating the period, phase, and amplitude of the two oscillatory motions. A difficulty that must be dealt with, however, is the nonlinear biomechanics of the limbs, to which Bernstein (1967) has called attention. It would seem that the translation process must involve a level that introduces offsetting nonlinearities in the motor-neuron signals (cf. Pew 1974). This may be thought of as tuning the system (Turvey 1977). The need to discover appropriate tunings may shed some light on the role of practice in the acquisition of motor skill. Euclidean maps and the brain. Humans using an external map (e.g., a road map) to navigate find it helpful to rotate the map until it and the world represented by it have the same alignment with respect to the navigator's body. When one is driving from San Francisco to Los Angeles, the map is held so that Los Angeles is out toward the windshield; on the way back, the map is rotated 180, putting San Francisco toward the windshield. This quirk highlights the fact that mapreferenced navigation requires the interrelation of three systems of coordinates - the world system, the map system, and the body system. I speculate that recent studies by Cooper & Shepard (1979) and Pinker & Kosslyn (1978) on the human capacity to rotate mental images of spatial configurations are examining a phylogenetically ancient process by which animals' maps are rotated into alignment with the environments they represent in order to make navigational computations simpler. The fascinating thing about the mental rotations examined in these experiments is that they are always and seemingly inescapably continuous. If the brain wishes to rotate the representation 180, it seems to have no choice but to do so by passing it through all the intervening angles. This suggests that the physical embodiment of spatial representations in the brain must be such that the only natural way to perform the rotation operation is continuously. The map that is embodied by a drawing on paper has this property. You cannot rotate it to a new orientation without passing it through the intermediate orientations. The matrix representations of spatial configuration in a computer memory, as simulated by Kosslyn, Pinker, Smith, & Schwartz (1979), do not have this property. In the matrix embodiment of a spatial representation, the rotation operation is carried out by computing a new location for the contents of each cell. There is no natural, unavoidable reason for computing intermediate locations. Indeed, such computations would seem counterfunctional in that they ought to increase the error in the final relocation. [See also Pylyshyn: "Computation and Cognition" BBS 3(1) 1980.] How then could representations of spatial configura-

The theory of action in cognitive psychology

Motor schema. Theorists outside the sensorimotor tradition have repeatedly suggested that learned movements, such as those made in writing, must be represented by "schemas," "movement formulae," and the like. This is another one of those cognitive notions that fill neurobiologists with unease. It seems so vague, and hopelessly remote from neurobiological embodiment. One of the things I try to do in chapter 12 is to give an explicit statement of what the neural embodiment of a movement schema could look like. A schema is an abstract representation of the form of a movement. It is posited to explain the fact that one's handwriting looks the same whether one writes at a desk or on a blackboard. Since the muscles used and the biomechanics are radically different in these two writing situations, the form of the movements made in writing must be specified independently of the motorneuron signals required to realize the movements. What one wants is a neurobiologically plausible way of specifying the trajectory of a movement without specifying patterns of motor-neuron activation. Yet there must be some way of deriving patterns of motorneuron activation from this specification of trajectory. Oscillators are a demonstrated reality in the nervous system, as is the principle of superimposing oscillator outputs. The Fourier theorem assures us that any trajectory may be represented as the superposition of the concurrent execution of - a suitably chosen set of oscillations. Anyone who has seen a chart of Lissajous figures will realize that limited modulations of the amplitudes, periods, and frequencies of two concurrent orthogonal oscillations can give rise to an astonishing variety of trajectories. These considerations make it attractive to assume that oscillations are the primitives in action schemata - the alphabet in which every schema is written. Thus, the schema for writing the letter "e" would consist of the specification of the required modulations in the periods, phases, and amplitudes of two orthogonal oscillations. At the time I made this extremely speculative suggestion I was unaware of the recently published work by Hollerbach (1981). I did not know how simple and elegant an oscillatory theory of handwriting could be. Nor did I know that recordings of movements made during handwriting and the nature of the distortions in rapid writing make such a theory empirically plausible. On the basis of Hollerbach's work, I am prepared to take the oscillator model of schemata more seriously.
618
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

Commen<ari//Gallistel: Organization of action tion be embodied in the brain in such a way that the continuity of the rotation operation becomes inescapable? One way would be for the perceptual system to decompose the to-be-represented configuration into its three-dimensional Fourier components (in effect, to represent the environment in terms of macroscopic matter waves, analogous to sound waves). The brain could then store the information about spatial configuration in the form of a set of five-dimensional signal vectors. One signal in each vector would specify the period of a Fourier component, another its amplitude, and the remaining three its phase and orientation. If information about spatial configuration were stored in this fashion, then the rotation operation would have the unavoidable continuity revealed by the mental-image experiments. The only natural way to rotate a spatial configuration represented in this way is to change the values of the signals within each vector that specify the phase and orientation of the component. Changes in the value of any brain signal are almost certainly continuous. The intermediate values that the phase and orientation signals would have to pass through en route to their new values would specify intermediate orientations. The exceedingly speculative character of this concluding suggestion hardly needs emphasizing. The suggestion is put forward only to illustrate a deeply held belief: behavioral studies, even studies of mental imagery, are an indispensable source of clues as to the primitives in terms of which the brain represents and acts upon the world. If we are to have neurobiological explanations of behavior, we must discover those primitives. ACKNOWLEDGMENTS I dedicate this precis to Donald Kennedy on the occasion of the celebration of hisfiftiethbirthday. Hefirststimulated my interest in the neural circuits subserving units of behavior in invertebrates and has greatly shaped my views. Also to Tony Deutsch, who first drew my interest to the nature of spatial representations and how they control action, and who set the course of my research in electrical self-stimulation of the brain. The expenses of preparing the precis were defrayed by NIH Grant NS 14935. Stelmach & Requin, 1980 for "the state of the art.") Particularly welcome are the reprints of papers by Sherrington (1906), van Hoist (1937), Wilson (1966), Fraenkel (1927), von Hoist & Mittelstaedt (1950), and Weiss (1941). The author's commentary will help the student better appreciate these classic contributions. If Gallistel had chosen the subtitle Classic perspectives I would rest content; but when a book tells me little that I did not know fifteen years ago, I find the claim of his subtitle, A new synthesis, somewhat suspect. Many ideas concerning "the organization of action" were introduced to the Western literature by Peter Greene in the sixties (e.g., 1964, 1967). Yet Green's ideas, far from being developed systematically by Gallistel, are dismissed in one sentence in a section entitled "The Turvey theory." Turvey is said to have been influenced by Bernstein and Easton, but (as I am sure Turvey would agree) Turvey s primary influence was Greene. Greene was one of thefirstAmerican workers to stress the centrality of the work done by the Russian school initiated by Bernstein, and Easton was Greene's Ph.D. student. I was also disappointed that the work of Bernstein himself, and the neurophysiological and robot locomotion studies of this "descendants" in Moscow, are only mentioned in passing. Another incredible omission in Gallistel's "new synthesis" is all mention of artificial intelligence, save for the erroneous ascription, on page 365, of the notion of heterarchy to Minsky & Papert (1972) rather than to McCulloch (1945). Rather than review this literature, let me simply cite two papers. Sacerdoti (1974) builds on a number of classic studies in robotics to study hierarchical planning. This work is approvingly cited by Shaffer (1980) in his study of the timing and sequencing of action. Again, Kuipers (1978) offers an Al approach to cognitive maps, while Arbib & Lieblich (1977) counter Gallistel's claim (p. 381) that "Deutsch's (I960) model for the control of action by a cognitive map . . . is . . . the only model that explicitly specifies a physically realizable process by which a motivational signal may organize the use of a map." Talking of cognitive maps: I would like to have seen a thoughtful discussion of the light an action-oriented approach sheds on the analysis of perception. True, much of a paragraph is devoted to the 1963 work of Held & Hein on the role of activity-induced sensory feedback in behavioral development, but this is too little too late (p. 378 of 394 pages of text). [See also Gyr et al.: "Motor-Sensory Feedback and Geometry of Visual Space" BBS 2(1) 1979.] One final nomination for my (distressingly incomplete) list of ingredients in a genuinely new synthesis: Piaget-inspired studies of the role of action in the cognitive development of the child. For an excellent review of this approach, see Forman (1981).

Open Peer Commentary

Commentaries submitted by the qualified professional readership of this journal will be considered for publication in a later issue as Continuing Commentary on this article.

On a clear day you can see behavior

Robert C. Bolles
Department of Psychology, University of Washington, Seattle, Wash. 98195

A new synthesis?
Michael A. Arbib
Computer and Information Science Department, University of Massachusetts, Amherst. Mass. 01003

Gallistel's The organization of action is a welcome book, and should accelerate the growing integration of the study of motor behavior into cognitive science and neuroscience. (See

By and large, psychologists have remained totally oblivious to the good work that has been done by other scientists on the organization of action. One reason for this curious neglect has been noted by Gallistel. Up until about a dozen years ago, the majority of psychologists, and surely the great bulk of psychologists who were concerned with learning, subscribed to the S-R formula. All responses, learned and unlearned alike, were controlled directly by their controlling stimuli. When learning occurred, it was the action system itself that became modified. Reinforcement and classical conditioning both resulted in the formation of S-R associations. Today, of course, hardly anyone subscribes to the S-R formula, because
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

619

Commentary/Gallistel: Organization of action

there is a new formula to believe in. Everyone (almost) still believes in associationism, and studies association formation, but the new associations are quite different: they are S-S connections. When learning occurs now, it is the perceptual system, or the memory system, or something of that sort that is modified. That is all well and good as far as the study of learning is concerned, but where does it leave our understanding of behavior? How do we explain the emergence of new behaviors when learning has occurred? For that matter, how are we supposed to explain any kind of behavior? How is the modern psychologist, caught up in the analysis of memory and the formation and organization of S-S associations, able to say anything about what the organism is doing? It would almost appear that when we abandoned the old S-R formula, we also turned away from behavior. The search for appropriate and realistic explanations of behavior has always been a serious business, but with the present involvment in S-S mechanisms the problem has become all the more serious. It is now an urgent matter. Gallistel has not only seen the problem and tried to start doing something about it: he has started off, I believe, in the right way. He starts with the outside of behavior, the oscillators, the reflexes, the automatisms, and encourages me to believe that he will ultimately be able to explain a lot of behavior with these devices. At least, the details of behavior will be explained at the level at which the detail is relevant. Why the behavior is occurring at all, is quite a different matter; that is the inside of behavior, of which I will speak in a moment. The key to this approach, of course, is the idea of hierarchical organization. We might also call it the principle of local autonomy. Gallistel illustrates the idea in a variety of ways, but I cannot resist another illustration: One day the word comes down from upstairs at General Motors, "We have a new goal: 300,000 Pontiacs this year." That is all they have to say, because the people at Pontiac Division know what to do. They get orders out to the factories, start stockpiling materials, get on the phone to Personnel to hire new people, contact the crew at the Sales Department, send a man over to the bank to get credit ready, and so on. Each of these groups knows how to do what it needs to do, and each of these groups knows how to command other people down the line to do what they have to do. When it is all over, about 300,000 people will have been involved in making and selling about 300,000 Pontiacs. The process is similar in complexity to an animal walking across the room. And it works. I am inclined to believe that it works for the same reason - namely, that such a complex operation can only be carried out if there is local autonomy all up and down the line. Interestingly, if we can believe Wright (1979), when something goes wrong at General Motors it is likely to be because the people upstairs have intervened in the process and tried to tell someone down the line how to do his job. Those of us who are highly encephalized and who can intervene in our own behavior can also testify how much more smoothly we move, and how much more likely we are to reach our goals, if we don't. My main complaint with the book was that it was far too short; it did not tell me nearly as much as I wanted to know about many topics. Time and again, just as I thought I was going to find an answer to some question that had been stirred up, the section or the chapter would end. It is clear that a great deal of work remains to be done, and that Gallistel has pointed the way to much of this work. I will note just two further topics that I think need much more examination. Motivation tends to be regarded as an inner phenomenon; surely the total switch of programs, or dispositions, that occurs when the animal suddenly becomes frightened or gradually becomes hungry must take place upstairs. But there is also local autonomy in motivation. Thus, while the general knows why the soldier is marching down the road (better than the soldier knows), only the soldier knows why he is smoking a cigarette. We have to think of motivation in the small as well as on a global or programmatic scale. Local control can be seen in the sequencing of movements. Thus, as the lad runs up to the soccer ball, we can see his intention to kick it. The body tilts back, and he takes the short step just as he approaches it. The inner sequence is: run up and kick it; the outer sequence includes a moment of intricately interwoven and overlapping motor coordinations. Gallistel has emphasized the important servo type of correction mechanism. But there is another, and just as important, kind of correction or compensation device. Consider this: when a man raises his hand to hold it out in front of him, there is a compensatory tilting back of the body at the hips. It appears that balance is restored by another one of those negative feedback loops. But that is not what actually happens. Much of the time, at least, the tilting of the body precedes the lifting of the hand. We have an anticipatory, or "feedforward" mechanism to make the correction. We have to learn a lot more about such devices. Indeed, we have to learn a lot more about all of this.

Independence and interaction in behavioral units

William Chappie
Biological Sciences Group. University of Connecticut, Storrs, Conn. 06268

Much of the experimental work on motor systems over the last one hundred years has been performed within a theoretical framework that viewed movement as composed of elementary unitary reflexes connected together in varying patterns. Gallistel's The organization of action represents an ambitious and partially successful attempt to update our notions about the way in which animal movements are generated by nervous systems. This is a good book, thoughtful and challenging in its attempt to come to grips with the central problems of movement. It does not represent a new theory, but rather a summary of where we are now. Yet perhaps it is a measure of the book's attempt to provide a reductionist theory of movement that I found myself questioning its premises at a number of points. Reflexes, oscillators, and servomechanisms represent the elementary units of movement. These "quanta" are assembled into more general sets which may in turn be associated with other sets of quantal units in hierarchical fashion. Potentiation or depotentiation of various levels of such a hierarchy can then produce the combinations of the elementary units required to generate a particular movement. Generation of movement in the external world often involves "cognitive maps" of that world. Portions of classic papers by Sherrington, von Hoist, Weiss, and others are used to present the essential elements of this theory. Perhaps most troubling is the notion of these elementary units. "Reflexes," "oscillators," and "servomechanisms" each describe an extraordinarily diverse group of phenomena. Some may indeed be elementary units; others may be composed of subunits. Is the oscillator of the scratch reflex an elementary unit evoked by locomotion as well as scratching or is it an intergral part of this reflex? Units observed in a "simplified" animal preparation may not be the same units employed in the normal animal. The alphabet of a class of movements, therefore, may not be divided naturally into reflexes, oscillators, and servomechanisms, but along other more diverse lines. There are, moreover, the difficulties of associating physiological units with behavioral ones. A discrete network of cells, such as the stomatogastric ganglion of lobsters, may exhibit a variety of different behaviors. Due to its nonlinear properties, a single cell may be a single physiological unit but several

620

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4)

Commentary/Gallistel: Organization of action

behavioral ones, acting as an oscillator and as part of a simple reflex loop. What constitutes an element in the hierarchy may depend upon the operations that define it. Thus, I do not wish to quarrel with Callistel's aim of constructing more general movements from elementary units, but simply to question how useful historically derived categories are in constructing a general theory of action. A second difficulty arises from the role of hierarchy in a theory of action. It seems to me that Gallistel has minimized the difficulties surrounding hierarchy. The existence of extensive interactions, both behavioral and physiological, between related but separable movements is found in many systems, from eye movements to locomotion. Gallistel himself admits this in discussing the concept of "relative hierarchies." Is it true that this lability is found only at the higher levels of the action hierarchy, or can this also be considered to be operating in the interaction between pacemakers in simple systems? Doesn't the notion of a fixed hierarchy depend upon the way in which feedback is used to control a particular movement? But these are arguments with a book that is, after all, on the side of the angels. Simplifications are useful if they help us on our way. By assembling a reductionist theory of action, Gallistel has shown us where we are, but also how far we have to go. to describe some of our recent observations on motor aftereffects. These aftereffects have been shown to be involuntary, patterned movements of limbs subsequent to treatment conditions which may or may not involve movement, but which normally involve direct or patterned sequences of effort (Craske 1981). Investigation of motor aftereffects in these laboratories has recently revealed a number of involuntary oscillatory phenomena, which provide good evidence for the presence of oscillators at a low level in the human motor system. We all know of the after-contraction effect, in which a subject first presses a wrist against a wall for about twenty seconds and then steps away, after which the arm rises involuntarily for a few seconds and then returns to the side. This commonly accepted decription is incomplete, however; the nature of the aftereffect is in fact a prolonged oscillation. It requries only that the subject stand bending forward with arms hanging down and that large involuntary oscillations of the treated arm occur. These oscillations involve as prime movers both the previous agonists and the previous antagonists; that is, the after-contraction effect is not restricted to the previously active muscle. The arm may sweep through 60 or more degrees, and for different subjects the period has ranged between 6.2 and 159 seconds. Furthermore, despite earlier reports, the duration of the phenomenon is between 5 and 20 minutes, rather than just a few seconds. If both arms are strained simultaneously, but in orthogonal directions, then, on cessation of strain, each arm will oscillate in the plane of its own previous effort. In this case there is a tendency for the frequency and phase of both oscillations to be strongly linked. In test conditions which follow both arms' being strained in the same direction, the aftereffects show more independence, and the arms may move with different amplitudes and frequencies. Duration of oscillation may also vary, one arm coming to rest while the other continues its involuntary movement. When one arm is strained in two directions following a sequence such as: forward for two seconds, brief relaxation, sideways for two seconds, and so on for twenty cycles, then when the arm is allowed to move of its own volition, oscillatory synthesis takes place. That is, the limb describes Lissajous movements (e.g., moves in circles or figures of eight) as the two orthogonal oscillations are released simultaneously and exhibit differences in the phase relationships and frequencies of the component movements. These combinatorial motor aftereffects have been observed for the eye and tongue, as well as the arm and leg. The fact that there is a phase shift between the two components in this situation indicates that the temporal relationships of the treatment are in some way being taken into account in the release of the aftereffect, giving a hint of hierarchical organization above the more basic oscillatory process. There higher levels of organization manifest themselves clearly when more complex effortful activities are used as treatments. For example, a subject is asked to pedal an excercise bicycle, or to shovel heavy snow or sand, or to stand on one leg and lift the other up and down while it is attached to a moderate weight. At the end of about one minute of this treatment, the subject lets the relevant limb or limbs hang free, and allows any involuntary activity to occur; after some delay, a very slow, involuntary, approximate facsimile movement will occur, showing the presence of strong organizational properties. In the case of the bicycle, however, the pedaling aftereffect can sometimes be in the direction opposite to the treatment, strongly supporting the involuntary nature of the aftereffect. The latter observation points to the presence of each of the components of the original action, but an absence of organization from within the motor system for this simpler situation. Thus oscillatory movement of the hip and knee joint are triggered at the wrong time - that is, with abnormal phase relationships. This brief summary of some recent observations serves, I
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

Oscillators in human motor systems

Brian Craske
Department of Psychology. Memorial University of Newfoundland, St. John's. Newfoundland, Canada A IB 3X9

The aspect of Professor Gallistel's book on which I want to focus attention is that of the role of the endogenous oscillator in motor behavior. I strongly echo his argument that psychologists have in general been slow to recognize the theoretical power of such a mechanism, and have made little attempt to exploit its interesting properties in their models of motor behavior. They have also been more than a little reluctant to acknowledge that there is a wealth of experimental evidence for the existance of these mechanisms in creatures as diverse as the cat and the cockroach (Delcomyn 1980). On of the main themes in the book must be read as a critique of those whose thinking is wedded to the reflex as the only functional unit of motor organization. Perhaps the time has now come, forty years after von Hoist's (1937) important paper, to explore the modus operandi of motor behavior in man without the self-imposed limitation of the chained reflex. Gallistel has reemphasized the likely importance of coupled-oscillator circuits as powerful and versatile components of motor systems, and he has ably marshalled the evidence to support such a view. I concur with his argument; it is this very versatility which points to the fundamental part which such circuits could play in the manifold actions which characterize so much of the behavior of higher animals. Perhaps one of the reasons psychologists have been tardy in the use of the oscillator concept is the virtual absence of readily distinguished oscillatory motor behavior in man. Although movements of a limb may be analysed to show that they are composed of oscillations in different planes, this does not show that oscillators themselves are involved in movement production. It would be valuable if evidence for the presence of relatively long-period oscillations in humans could be presented, for if limbs were shown to exhibit these movements involuntarily, then the oscillator must at least be considered as a potential underlying mechanism. In order to provide empirical support for the argument concerning the existence of the endogenous oscillator in humans, I would like

621

Commentary/Gallistel: Organization of action

hope, to add some impetus to one part of the explanation of motor behavior espoused by Gallistel, for these data can be explained most parsimoniously by invoking the oscillator as an underlying mechanism in human motor behavior. feedback. Together these commands specify a reference trajectory and an error criterion. Is the reference trajectory the same as the servomechanism? The reference trajectory that drives the robot arm generally operates in the servo mode, but the specification of the reference trajectory is based on knowledge of arm mechanics. The reference trajectory in the animal also may operate as a servo if the error criterion is used. However, afferent feedback is not necessary for the production of a transient movement, as has been shown for head movements in the deafferented animal (Bizzi et al. 1976). The specification of the reference trajectory itself is done open-loop, presumably based on an internal representation of limb dynamics. Thus there appears to be the need to include, in a theory of action, a non-oscillator-based programming unit that can specify the transient characteristics of a movement and that can drive a movement irrespective of the presence of feedback.

The reference trajectory as an organising principle

Parvati Oev
Rehabilitative Engineering Research and Development Center, V.A. Medical Center, Palo Alto, Calif. 94304

Gallistel has done a good job of putting together a complex set of concepts and synthesizing a theory of action. As generally does happen in such an encyclopedic work, some interesting aspects of the field fall through the cracks. One such interesting aspect is the generation and control of transient movements. The history of the movement field has shown us the basic nature of the reflex, the servomechanism, and the oscillator as organising principles. Transient movements, such as reaching and turning to a stimulus, appear to be preprogrammed rather than servo-operated, and do not seem to use the reflex or the oscillator as their basic unit. What are the organising principles underlying transient movements? Is Fourier theory, as suggested by Gallistel, a possible organising principle? In the strict sense, any time signal, such as joint position during movement, can be decomposed into a sum of sinusoids, the Fourier components. Though this may be an economical description for a periodic movement, for a transient movement the set of frequencies or Fourier components is infinite. Therefore, there is no simplification of the control problem if one attempts to generate the Fourier components rather than the time signal of the movement itself. Control of transient movement has been studied for the movement of robot arms. The desired movement of a robot arm is specified in terms of a reference trajectory, generally the trajectory of desired joint angular velocities and specified intermediate and final points. Bizzi et al. (1976) have proposed that the nervous system specifies those activity levels in agonist and antagonist muscles that are required to maintain the desired final position. Elasticity of muscle then ensures that the limb, or in this case the head, moves to the equilibrium position. This corresponds to specifying a reference trajectory where the trajectory is a step change in reference position. Ghez & Vicario (1978) show that, in an isometric force-development task, more motor units are activated during the initial phase of rapid force development than when steady force is maintained. In other words, the reference trajectory may consist of two signals - an early pulse specifying the initial rate of change of force, and a maintained value specifying the desired final force. Robinson (1973) noted that the pulse-step from the signal to the muscles was suited to the acceleration of a mass followed by maintenance of a position against an elastic restoring force. Thus one can conceive of a single movement command's being translated into a set of subcommands, one specifying the desired final position, another specifying transient characteristics (such as the peak velocity, perhaps as a function of the level of potentiation of a motoneuronal pool), a third specifying how rigid should be the adherence to the desired trajectory in the face of unexpected disturbances, and so on. Each of these commands may be transmitted to different neuronal pools - the command for desired final position to a pool of small motoneurons, the specification of peak velocity to a pool of large motoneurons, the command regarding trajectory precision to neuronal pools monitoring afferent

Nodes, notions and neuroscience

Robert W. Doty
Center for Brain Research, University of Rochester Medical Center, Rochester. N.Y. 14642

Gallistel's book is, unequivocally, an important and creative analysis of the manner in which the central nervous system coordinates and initiates purposeful movement in creatures as diverse as insects and primates. Most impressive is the author's ability to extrapolate exceedingly complex action from a few underlying, relatively simple principles. The examples reproduced in extenso from the original literature are well chosen and lucidly explicated, particularly the principle of the "efference copy." Gallistel's emphasis on the lattice nature of hierarchical organization in the nervous system is perceptive and useful. Perhaps the major contribution will turn out to be the suggestion that afferent activity is encoded into patterns appropriate for organized output. While no one can currently say what this means, it provides a reasonable guilding hypothesis in a field of endeavor still struggling to identify a meaningful code in the digitally transmitted representations of sensory events. It may also provide a new insight into the manner in which the afferent portal "gates" the input into particular channels and thereby selects which of many reactions will occur to pluripotential input (Doty 1976). The "gating" is most probably based on the spatiotemporal pattern of the afferent discharge, but, in view of Gallistel's penchant for "oscillators," it bears noting that, in one instance where "oscillation" and "resonance" could reaonsably have been expected, it was almost certainly lacking in the central neural mechanism (Doty 1951). The book is sometimes annoyingly pedagogical. It is hard to imagine a reader who would stay with the exposition to page 299 yet not know what an electrode is - but it is duly explained to him. More important than annoyance, however, this unnecessarily constrained approach limits the documentation of many of the facts alluded to, and the reader is carefully spared many of the complications which subsequent research has uncovered - e.g., on the "myotypic response." Although inhibition as a mechanism of coordination and control is clearly identified by Gallistel, I think its importance, both in restraining the operation of irrelevant synergies (Roberts 1974; Doty 1976) and in accessing motoneurons for ensuing patterned activity (Doty 1976) is not sufficiently emphasized. Indeed, Pearson and his colleagues (1980) have

622

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4)

Commentary/GaWisteh Organization of action

now elegantly detailed the manner in which visual input causes a grasshopper to jump; and inhibition is a crucial factor! As Gallistel's discussion wanders farther from neurophysiological principles one gets sentences like the following: "The assumption that food nodes on the animal's cognitive maps are likewise connected to this hunger center means that a potentiating signal from the hunger center is automatically conducted to nodes representing food" (p. 352). To his credit, Callistel recognizes this as merely nominalizing a problem; yet this is how Weiss and Skinner "went wrong" - by simply verbalizing, describing, or hypothesizing processes without reference to their neurophysiological likelihood. On the other hand, one must eschew pretense at understanding that which remains basically mysterious. The "nodes," the (digitalized) Fourier transforms, and the preferential "potentiations" carry, in essence, no more explanatory power than the "homunculus," or the "psychical" activity of which Sherrington wrote so poetically, and which Callistel so uncategorically discredits. Forthright identification of the limits of knowledge is an essential feature of its expansion - no Copernicus so long as all epicycles are believed to turn in perfect coordination. The ultimate understanding of the "organizaiton of action" must include mnemonic and conscious contributions, and, for what is presently known about the nature of these processes, the critical feature of either "homunculus" or "node" is the profession of ignorance. Callistel confesses that he has a certain "feeling" when his cognitive map comes into congruence with reality. One wonders whether the wasp does likewise - an idle question, were one not able to hope for elucidation of the underlying neural processes in each case (Doty 1975; Bartlett & Doty 1980). tively altered. More recently, at a more restricted level of analysis, a number of workers have shown that even relatively simple motor acts (such as scratching in turtles; Stein & Grossman 1980) can often be fractionated (e.g., into incomplete cycles plus variations of muscular action from one cycle to the next). Similarly, Colani (1976) demonstrates how a particular action in mammals can be performed reliably through a variety of kinematic articulations. The unitary nature of "units," whether defined at motivational or at motor levels, is thus, at best, a useful approximation. 2. Are behavioral units separable (I.e., Independent)? The force of this question was first brought home to me when I found that a variety of independently defined motivational systems (e.g., "fear" or "exploration") could facilitate, as well as block, rodent grooming actions - partially as a function of the degree to which these apparently unrelated systems were activated (e.g., Fentress 1976). It was brought home a second time when I found that the detailed form of an individual grooming stroke could be influenced by its neighbors (Fentress 1976; 1980; in press), a phenomenon not that different formally from issues of coarticulation and perception of the "segments" of human speech (e.g., Studdert-Kennedy 1976; Liberman 1980). A third lesson came to me from the elegant descriptive study by Woolridge (1975) on grooming patterns in mice. His detailed measures of area of the face covered by the paws, velocity of individual movements, etc. during mouse grooming sequences all seemed to reflect a continuous distribution of processes that crossed over the categories of action that I had previously proposed (e.g., Fentress 1972). A final example can be seen in the conceptually important unpublished dissertation by Bellman (1979), who showed that individual features of feeding and fighting in lizards may be blended, given the appropriate circumstances. In short, "units" of action, as we define them today, are not immutable or fully separable, but can vary within limits, and even overlap, in ways not yet fully understood - but in some sense as a function of the broader surroundings within which they, as individually defined, occur. 3. Are behavioral units fixed (I.e., static)? It would certainly be nice if they were, but I doubt whether this will be found to be the case. Behavior, after all, is fundamentally a nested set of more or less coherent processes, not a set of "things" in any strict sense. Let us take the idea of central pattern generators (which Callistel reviews admirably) as an example. By now we should expect that pattern generators are neither unitary entities nor necessarily independent, in a strict sense, from their surround (e.g., Selverston 1980, and accompanying commentaries). The only points I wish to emphasize here are that the balance between central regulation and sensory influence can vary systematically as a function of movement, speed, etc., and that even the specificity we assign to a behavioral control system need not be fixed. Thus, Woolridge and I have recently found (unpublished) that the degree of influence of specified extrinsic factors upon the patterning of grooming actions in mice can shift systematically with the speed at which these movements are performed (cf. Fentress 1980) - as if "the central program" became more autonomous from the periphery when it was strongly activated. This implies that the boundaries of control processes in any given form of behavior (i.e., the defining characteristics of a "unit") are modifiable, but in a rule-given way. Analogously, the "breadth" (specificity) of control boundaries that participate in grooming and a variety of other specifically defined actions can vary systematically as a function of stimulus-strength, time over which a set of stimuli operate, etc. Many other examples could be given if space limitations permitted. Toward a view of coherent processes as an alternative to unit structure. So, what does all this mean with respect to Gallistel's helpful attempt at providing us with a synthesis of
THE BEHAVIORAL AND BRAIN SCIENCES (1981). 4

Network foci in integrated action: Units or something else?

John C. Fentress
Department of Psychology, Dalhousie University. Halifax, Nova Scota B3H 4J1

As Callistel emphasizes throughout his valuable text, a major challenge to the behavioral and brain sciences is to examine the network of events that (a) permit the organism to articulate distinguishable actions and (b) mold these actions together into functionally coherent ensembles. In addition, as the author emphasizes, it is essential to seek rules by which the various "levels" of integrated performance (e.g., movement and motivation) may be conceptually joined together a task that I too have been especially concerned with in recent years (Fentress 1976; 1980; in press). Callistel has understandably chosen to base his arguments on a unit-structure model of behavior and the hierarchical arrangement of these separately defined units. It is upon these points that I focus my comments. What is a unit of behavior? Dictionary definitions of the term "unit" lead us to anticipate a structure that is (a) indivisible, (b) separate, and (c) stable. It is useful to explore each of these attributes further.
1. Are behavioral units Indivisible (I.e., unitary)? This

depends upon how and where we look. My own work on rodent grooming, to which Callistel kindly refers, shows that individually defined "actions" have a statistical, but not certain, association to one another in time. It is, for example, possible to perturb some features of grooming while leaving the others quite unaffected; i.e., the grooming unit is subdivisible in its rules of expression. At a higher level of organization Hinde (1959) noted some years ago that motivational categories ("drives") can be fractionated - e.g., certain components of feeling or fighting or mating can be selec-

623

Commentary/GaMisteh

Organization of action

integrated action? I think it means that our definitions of "units" of behavior must be rethought. I think it means that units are, to borrow from Sherrington (as quoted by Gallistel), "a convenient . . . fiction" - useful, within limits, until something better comes along. And what form will this "something better" take? Here I can only speculate, though I do so with reason. It will take, first, an explicit awareness that integrated behavior is a nested set of more or less coherent processes rather than a set of indivisible, independent, and separate things. It will take an increased appreciation for the rules of relation amongst separately defined processes as a key to understanding (in a deeper sense) these processes rather than merely the reverse (i.e., the assemblage of individual processes to construct higher orders of organization). That is, it will take an improved awareness amongst the neurobiological community that change and relation are key ingredients in their enterprise. A brief illustration here may help make the point. One basic feature of many action systems is that the subprocesses of movement may be varied over a wide range while coherence at a higher level is maintained I can brush my teeth by moving my arm or my head, or innumerable combinations of the two; a mouse can make sweeping contact with one part of its face through a variety of kinematic combinations of adjacent limb segments, head and forelimb movements, etc. Coherent processes at one level - either "higher" or "lower" may be reflected in variations at another. One thus frequently observes what I have previously called relational constancies in behavior, where coherences in "higher levels" of expression (toothbrush bristles passing over the teeth) are maintained through complex variations at "lower levels" of expression (individually defined limb movements; cf. Golani 1976). It is not one level versus another, but some synthesis among them that can itself vary in time. In some respects, as Gallistel emphasizes, future research will need to exhibit an increase awareness of problems in relating hierarchically distinguished levels of behavioral (and/or brain) organization. I do not think, however, that it will simply be a case of the foci of network actions being examined at a variety of levels of order, or of one being used to explain another (cf. Rose 1981). One can but hope that each of these levels will contribute something of significance to its neighbors - not as "units," but as more or less coherent processes in their own right. This is my own extrapolation of the important message of Gallistel's book. While I might appear to quibble with him on certain details of exposition especially his treatment of "units" and unidirectional "hierarchies" - I applaud his efforts to lead us toward a more integrated appraisal of integrated function than was otherwise readily available. Summary. So, let me summarize briefly: "Units" of behavioral action are sometimes convenient fictions. They provide a clear format for distinguishing one feature of behavior from another, a task, as Uttal (1978) notes, that is much more difficult than is the separation, say, of one anatomical (e.g., brain) locus from another. Behavioral "loci" must take into account not only that behavior is multilayered, but that at each layer behavior involves constructs of change (i.e., process), and that both between and within layers there are a multiplicity of nonfixed relations. What we are left with, then, is the search for foci in both the expression and the control of integrated action in a manner that is appropriately fluid as well as multilayered in its content. Models that are truly dynamic, relational, and multilayered will, in the future, I suspect, make unit-structure and linear-hierarchy conceptions of behavioral and brain science seem as unsatisfying as are static, isolated, and linearly ordered mechanical concepts when viewed in the context of modern physics (e.g., Bohm 1969).

"A new synthesis?"

Sten Grillner
Department of Physiology III. Karolinska Institutet, S-114 33 Stockholm, Sweden

Gallistel's The organization of action: A new synthesis is for the most part enjoyable reading. To illustrate well-established, even if not always well-known, findings, Gallistel has chosen to print full papers or parts thereof (45% of the text in the book) from the works of Sherrinton, von Hoist, D. Wilson, Fraenkel, and Weiss. These classic papers are provided with comments relating them to more recent findings. In this way Gallistel deals with the Sherringtonian reflex concept, coupled oscillators, servomechanisms and efference copy, and hierarchical organization of movement control - which are all placed in relation to animal behaviour - for instance, central motive states. He arrives at a very general picture of different aspects of movement control (action) that, as far as I can tell, is not in the least different from the one that most of my colleagues in my own field have had for years, and that has guided much of the work in this field at least throughout the seventies. There is nothing scientifically new in this book. It is therefore difficult to provide a commentary on what is presumed to be "a new synthesis." Part of this field has been a subject of discussion in some depth in prior BBS target articles by Selverston (1980) and Kupfermann & Weiss (1978). What is new is that a physiological psychologist, active in circles with little knowledge about the neural control of movement, has taken the trouble to go through this literature. He has been pleasantly surprised to find that in fact a great deal is known. He has therefore written a book to inform his colleagues and students about his new treasure. This is all very good, and by reading this book many researchers may certainly add to their perspectives. To call the book "a new synthesis" is, however, not only presumptuous, but in fact scientifically misleading. One might suspect that the title was imposed on the author just to improve the sales of the book (I bought the book for that reason myself), but this does not appear to be the case on the basis of what Gallistel himself writes about his conceptualization. (The same unfortunately appears to apply to another "action theorist," Turvey 1977.) The book reads well, and I did not discover any major errors. It takes examples from a limited group of motor acts, frequently locomotion, in a variety of animals. To my mind it would seem essential at least to go through all the main types of movements in a systematic and penetrating fashion rather than to extrapolate from a few examples. It is clearly too much to ask to expect that the author know this wide literature well, since he is not in the field. Therefore it must be accepted that quotations, references, and examples will seem rather arbitrary and the account of different phenomena sometimes superficial. What I have just said may seem harsh. I would like to add, however, that is is a most admirable feat to write this book on "action" (motor control) from the outside in order to bring knowledge of principles of motor control to the field of cognitive psychology. We are now in an exciting era when the boundaries between different neurobiological and psychological disciplines break down. This book is a manifestation of this trend: what can be more important than to unite cognitive psychology with basic neurobiology? This was expressed by von Hoist in the following way: "It is hoped that this article will contribute to the gradual disappearance of attempts to describe the functions of the highest developed organ of the body with a few primitive expressions. The sooner we recognize the fact that the complex higher functional Gestalts which leave the reflex physiologist dumbfounded in fact send roots down to the simplest basal functions of the CNS, the sooner we shall see that the

624

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

Commentary/Gallistel: Organization of action

previously terminologically insurmountable barrier between the lower levels of neurophysiology and higher behavioral theory simply dissolves away" (von Hoist Mittelstaedt 1950, and p. 209 in Gallistel 1980). What I would like to see, however, is a new book by Gallistel or somebody of his calibre, on the same general theme, provided that he had specialized coauthors for each of the chapters who could provide the necessary precision and nuances for all statements made. Such a book could become a major contribution to science. The present book is useful as an introductory text for graduate students. hierarchical organization, calls attention to a distinction between hierarchies of connection and hierarchies of classification or embedment. In a hierarchy of connection, superior elements exert control over inferior elements in the same way in which a captain gives orders to a lieutenant. In a hierarchy of classification or embedment, inferior elements are actually a part of superior elements in the same way in which a platoon is part of a company. Platoons and companies are different classifications of the same elements: soldiers. If we apply this distinction to Gallistel's hierarchy, we see immediately that at least the first five levels posited by Weiss are a hierarchy of classification With respect to neural connections, thesefivelevels are in fact at the same level - the level of the primary motor neuron or "final common path" of Sherrington. The primary motor neurons are of course interconnected and are subject to mechanisms such as reciprocal inhibition, but there is no evidence to suggest that these interconnections are hierarchically organized. Actually, Weiss's whole system is a hierarchy of classification in which behavior in the service of the animal as a whole is logically the highest level of organization: Weiss never implied a connective hierarchy of the sort Gallistel attempts to construct. In a sense, Gallistel does recognize this distinction, because he finds it necessary to postulate "intermediate units of organization" (p. 288). Intermediate units have a degree of complexity that would put them at about level 5 in Weiss's system. I would argue that these intermediate units are the lowest level of behavioral organization in terms of a neural hierarchy, and that Weiss's lower levels are merely embedded in these units. Once we ascend from the level of an intermediate unit, I think, the evidence for hierarchical neural organization is unassailable. Gallistel's presentation and discussion of the evidence are clear, concise, and convincing. It is perhaps worth mentioning, however, that hierarchically organized behavior systems are not the only way to activate motor neurons. There are several parallel routes, including the well-known pyramidal cells from the motor cortex. It seems certain that these motor-cortex cells are not generally controlled by diencephalic neurons. Such influences from different parts of the nervous system can be superimposed at the level of the primary motor neuron, but are not themselves part of the behavior-system hierarchy. Finally, let me mention the problem of levels of analysis. My point here is implicit in what I have just said about intermediate units. It may be possible to describe the basic elements out of which action is organized as reflexes, oscillators, and servomechanisms. But even the simplest spinal reflex involves coordinated movement of a whole limb, and single oscillators are postulated to control movements of whole fins in fish by von Hoist (1937) and of whole legs in insects by Wilson (1966). Thus, these elements have a degree of complexity somewhere about Weiss's (1941) level 5. Organization of action at a lower degree of complexity may not be behaviorally meaningful. In conclusion, I think Gallistel has done an excellent job of presenting a theory of action in a stimulating and provocative way. Consideration of the sorts of points I have made should pique discussion that may lead to clarification and refinement of the basic theory.

Hierarchy and behavior

Jerry A. Hogan
Department of Psychology. University of Toronto, Toronto, Ontario, Canada M5S 1A1

The organization of action is an interesting, intelligently written, stimulating, and provocative book. It is interesting because of its format. It reprints, in whole or in part, six classical papers on the neurophysiological mechanisms underlying action patterns. Gallistel precedes or follows each paper with commentary, and then integrates this material into a theory of how action is organized. The book is intelligent because the choice of papers is so clever that the reader is almost forced to embrace the theory before it is formulated. It is stimulating because the original papers as well as Gallistel's own contributions sparkle with the enthusiasm inherent in presenting good ideas that are well supported. And it is provocative because some of Gallistel's speculations, in chapters 11 and 12, about how cognitive preocesses control action are truly outrageous. The goal of the book is to discover the units out of which action is formed and the principle by which the units are integrated. The units turn out to be reflexes, oscillators, and servomechanisms, and the organizing principle is the hierarchy. In very important ways, Gallistel's theory of action is similar to Tinbergen's (1951) familiar hierarchical organization of instincts. Tinbergen's formulation was based on many of the same papers reprinted in The organization of action, so the similarity should not be surprising. What Gallistel has added is a more extensive discussion of the units themselves, a more satisfactory foundation of the theory in the facts of neurophysiology, and an extension of the theory to the realm of cognition. I shall comment on two closely related aspects of the basic theory in which I think some confusion exists: levels of analysis and hierarchical structure. Let me begin with the problem of hierarchy. Gallistel's argument for a hierarchical organization of action begins with the hierarchical system proposed by Paul Weiss (1941, partially reprinted in chap. 8). This system recognizes six levels of motor organization: (1) movement of muscle fibers controlled by a single motor neuron; (2) movement of a whole muscle; (3) movement of a single joint; (4) movement of a whole limb; (5) a locomotor act such as walking, jumping, or swimming; and (6) motor acts in "the service of the animal as a whole under the control of the sensory apparatus." Gallistel than proceeds to subdivide Weiss's level 6 into yet more levels. At least two such levels seem well supported: the level of the complex motor act (fixed action pattern) and the level of functional integration of motor acts (behavioral or motivational systems such as hunger, aggression, and sex). Now, throughout his discussion Gallistel implies (as does Tinbergen 1951, Fig. 98, p. 125) that this hierarchical scheme reflects "the organization of the underlying neurophysiology and neuroanatomy" (p. 292). Here, I think, is where the confusion arises. Dawkins (1976), in a recent descussion of

Effective procedures versus elementary units of behavior

John M. Hollerbach
Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, Mass. 02139

The main thrust of Gallistel's The organization of action is a

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

625

Commentary/GaUisteh Organization of action

broad attack on dualism as it pertains to the control of movement. By proposing specific mechanisms which can realize intentions, the author seeks to bridge the gap between cognitive intention and the actual production of movement. The author is addressing his arguments both to physiologists and to cognitive psychologists, with the ultimate hope of showing that some of the distinctions between the fields are only apparent, not real. Be defining cognition as the domain of mental representations, the author shows that even the lowly wasp must have cognitive abilities, due to its terrainmatching abilities. Without actually coming out and saying it, the author is using a well-known thesis in the theory of computation to justify the attempt to emody mentalist functions in physically realizable neural circuits. Church's thesis states that anything which can be formulated as an effective procedure can be computed by a Turing machine. Basically, what the thesis is saying is that any idea which can be described in sufficient detail can be physically realized. Church's thesis, it must be emphasized, is a hypothesis and not a proof. Nevertheless, the alternative is not palatable, as it takes us into the realm of mysticism, because there would be mental processes which we could not know and could not describe. Another line of thinking which supports the author's main thrust is the recognition that the central nervous system (CNS) had to evolve in the context of controlling its physical machinery. Animals have been propelling masses under the influence of gravity for eons, and somehow a representation of effective control of limbs must have worked its way into the CNS. The structure of the book is a series of introductions and discussions centered around excerpts from classic papers in motor control. The author emphasizes that he doesn't want a book of the latest word, and that the student of motor control is best advised to read classic papers. It is, of course, the author's prerogative to structure his book as he sees fit, but it is my opinion that this structure does not always facilitate the professed goal of convincing psychologists of the potential for physical embodiment of mental processes. Through the choice of such classic papers, the book takes on the tone of a history of ideas concerning the evolution of motor control. This tone sometimes interferes with the professed goal because the author must spend a great deal of time updating or disavowing some of the older ideas. Despite the author's reluctance to include the "latest word," I feel it would have been more forceful to start from the most current thinking about some of these motor-control ideas. The author has identified three principles which he believes are among the fundamental building blocks of movement - namely reflexes, servomechanisms, and oscillators. Moreover, he argues for a hierarchical structure for motor control, a paradigm used to describe many other aspects of CNS organization and hence readily acceptable to psychologists. The trump card for showing how intention can be linked to action is selective potentiation, which operates on the hierarchical motor-control structures. The papers the author has chosen provide experimental evidence to support his proposals. My main criticism of this choice of principles of motor control is probably one the author would readily agree to, but one which should be stated anyhow. These principles are coarse descriptions of some aspects of motor behavior, and are not detailed enough and probably not rich enough to constitute effective procedures for motor-control functions. The difficult question of what is an elementary unit of behavior cannot be answered yet, and it may be that the question is ill posed. The author has the greatest difficulty in making the reflex an effective concept. What makes a reflex particularly awkward for the author is that Sherrington was a dualist. The concept of a nervous system reacting to the world rather than 626
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

acting spontaneously is anathema to the notion of a central representation. By the time the author has circumscribed some of Sherrington's ideas such as chaining and a refractory period, and by the time he has described how the central nervous system can modulate reflex response, one is left wondering what the concept of reflex really means. The author seems to be incorporating more centralnervous-system activity into the concept of reflexes than is normally accepted. The role of reflex activity in the control of movement is a controversial and ongoing subject of research, and it is probably premature to elevate reflex activity to the role of a fundamental building block without further understanding of the nature of its contribution to motor control. In fact, many recent experiments on load compensation during arm movement in primates indicate that the gain of the spinal reflex is very low and that long-loop activity via the brainstem and higher centers provides by far the greatest part of load compensation. Some workers call this activity long-loop reflexes, but in so doing the term "reflex" has become just a label which doesn't provide any insight. Sherrington, of course, made significant contributions at the cellular level of understanding. Callistel also makes a point of diagramming specific neural circuits that could implement some of the processes he is illustrating, with an open admission of the potential problems this could lead to. The zeal to prove that intelligent processes can have a physical embodiment leads to the potential problem that too strong a concentration on the cellular level may miss the point of the computation. Since the author is not averse to using computer metaphors, focusing on cells is like trying to understand how a computer program works by measuring transistor outputs. Naturally the analogy to computers must be tempered with the knowledge that there is more locality to neural circuitry than there is to computer circuitry and that cells may be carrying out a higher level computation than transistors. Nevertheless, the recurrent inhibition network discussed by the author sounds like an XOR gate. An XOR gate is a fundamental logic unit in a digital circuit, but in and of itself it implies nothing about its use. Fourier theories of visual image processing and of auditory processing of speech, for both biological and computer-based systems, have been common. It was inevitable that someone would propose a Fourier representation of motor commands. Fourier representations have proven to be bad ideas in vision and audition, and it is distressing to see history starting to repeat itself now in motor control. A Fourier decomposition of a signal is often merely an exercise in curve fitting, and the curve fitting does not necessarily capture the essence of the signal structure. The term "servomechanism" is borrowed from control theory, where the meaning of this term is fairly definite. A servomechanism comprises a reference signal, a sensor for detecting the amount of error between the reference signal and the actual state of the device, and a feedback law which is used to synthesize new control signals based on the degree of error. This concept can be broadly applied as a paradigm to many motor activities, since animals often pursue some goal and modify their actions if they start to drift from the goal. Once again, however, there are problems in broadening a more specific concept. It is important to attempt to distinguish servomechanisms from adaptive controllers. The more logic that is involved in responding to an error signal, the more a controller becomes an adaptive one and less a servomechanism. In talking about wasp flight on page 148, the author remarks that the wasp taxis involves perceptual mechanisms of a high order as well as a memory. When there is this much logic involved in error correction, it is not clear what the ultimate error signal is, or how it is processed. Another problem with the servomechanism concept is the normally slow speed of conduction of nerve impulses relative

Commentary/Gallistel: Organization of action

to the speed of movement. Control theory tells us that servomechanisms based on error signals with significant delays will result in unstable systems. Intuitively, if it takes a long time to process an error signal and devise a corrective response, then by the time the response is operative the system will already have moved to a new state for which the response is inappropriate. The muscle spindle system would appear to be an obvious candidate for part of a servomechanism. However, long-loop delays from muscle to higher centers and back are on the order of 70-100 milliseconds, which, at least for the more rapid movements, is too slow to act as a servomechanism. Recent trends in motor-control research emphasize feedforward planning rather than feedback correction. Postural adjustments, particularly in locomotion, are made before movement commences or contact with the environment occurs. On page 138 the author explains how a servomechanism might interact with an oscillator for terrain adaptation during locomotion of insects. In higher animals, at least, terrain adaptation appears to have a greater cognitive or feedforward component: there is active anticipation going on. [See also Roland: "Sensory Feedback to the Cerebral Cortex During Voluntary Movement in Man" BBS 1 (1) 1978.] The role of proprioceptors in the control of movement is in fact quite puzzling. The more obvious hypotheses about the use of muscle spindles, Golgi tendon organs, and joint receptors, such as the servomechanism hypothesis, do not appear to hold. As good a guess as any at the present time is that proprioceptors signal unusually large deviations from intended movement, and that higher centers use this information to devise rather gross corrective movements. Along the line of differences in control systems between insects and higher animals, I feel the author should have directly addressed a point raised by Weiss (1941) on page 234. This point is that the higher the animal on the evolutionary scale, the more independent a control of joints and musculature it has. Consequently the type of control strucutre for movement may change along the phylogenetic scale. Anatomical and physiological observations strongly suggest some sort of hierarchical motor-control organization, as advocated in the excerpt from Weiss. The author modifies the notion of a strict hierarchy to a lattice hierarchy. In the limit, a latticed hierarchy becomes a heterarchy, where the concept of level disappears and every module may freely interact with every other module. Selective potentiation acting on a latticed hierarchy probably has the effect of turning the hierarchy into a heterarchy. The author makes much of the ability to set a general tendency through selective potentiation, which can result in a general course of action without a detailed specification of the steps needed to achieve the action. Setting activation thresholds for modules of a hierarchy and allowing interactions with the environment or other modules to trigger activation of a module is a well-known concept in the artifical intelligence (AI) field, where this sort of control structure is variously named production systems, daemons, actors, etc. This control structure is essentially heterarchical because the interactions between the modules are ill defined and are subject to local recognition of a pattern which results in activation of the module. Hierarchical versus heterarchical organization of control was a hotly debated issue ten years ago in AI. Eventually this issue was abandoned because the concepts of hierarchy and heterarchy were considered too simplistic to explain complex intelligent processes. To be sure, there are aspects of these processes which seem approximately hierarchical or heterarchical, but this is too rough a cut. Instead of applying such global paradigms as hierarchy, researchers in AI have focused on more specific problems such as language parsing or lowlevel vision and have let the requirements and constraints of solutions to a specific problem dictate the nature of the control structure. What usually emerges is a rich and unique control structure specific to the problem, and this control structure goes far beyond concepts such as hierarchy. In conclusion, the main concepts advocated in this book have had a historical impact on molding thinking about motor control, and this book is a useful history of the evolution of these concepts. The attempt at bridging the gap between physiology and psychology and at counteracting mentalist philosophies has been convincing. Whether these concepts constitute elementary units of behavior or control structure is, however, problematic, because the concepts are sometimes not clearly defined or are coarse descriptions of much more complex processes. Von Hoist's words on page 209 provide a fitting summation: "It is hoped that this article will contribute to the gradual disappearance of attempts to describe the functions of the highest developed organ of the body with a few primitive expressions."

Gems set into a base matrix

Rudolf Jander
Division of Biological Sciences, University of Kansas, Lawrence, Kans. 66045

Surely a new synthesis of how action is organized is what we desperately need. How promising is the title of Gallistel's book, and how thoroughly disappointing the text, when approached with the above expectation. Indeed, the book is "plagiarism on a grand scale," as the preface's first sentence apologetically admits; and thereafter one discovers that much of the text merely consists of reprints, albeit well-selected early gems. These are then set into a matrix of the author's own concepts and theories, which together constitute the book's original component, which will be scrutinized in the rest of this commentary. Given the blatant mismatch between title and text, doubts about single authorship of the two seem at first to be justified. However, after one finds how riddled with inconsistencies the original text is, this hypothesis becomes untenable. Since exposing and correcting all the flaws of logic and argumentation would amount to writing a whole new book, I instead pick out a few representative examples, first at the conceptual, then at the propositional level. Gallistel's logical starting point for his line of reasoning is Sherrington's idea that all complex behaviors can be decomposed into elementary units of behavior. Note that "elementary units of action" would better fit the context because this term appropriately excludes the behavioral-input side: sensory reception, perception, and cognition. After this minor blemish the conceptual muddle keeps building up, step by step. Without supporting reasons we are told that there are just three important types of elementary behavioral units; to Sherrington's one original type, the reflex, the author adds the oscillator, his only genuine elementary unit of action, and the servomechanism, which is rather a unit of behavior than a unit of action. These inconsistencies are topped by violating virtually all logical rules for constructing a scientifically useful categorical system. Thus, and most importantly, the elementary classifying concepts are neither mutually exclusive nor exhaustive, as they certainly should be. Notice that no objective criterion to speak of is given for drawing the line between reflexes and servomechanisms. Virtually all reflexes, as is well known, are themselves servomechanisms, or at least components of servomechanisms. In order to escape from this artificially created conceptual dilemma nothing but a subjective solution is offered: the reflex - as compared to a servomechanism - is a servomechanism whose feedback component is (arbitrarily?) not considered (p. 11).
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

627

Commentary/Gallistel:

Organization of action facts pales in comparison to workability; even negative data mentioned at other points in the book are ignored, for example, nonlinear summation (p. 41 ff.) and the magnet effect (p. 99ff.).My main argument in this context would be that if some behavioral pattern is realized, this is in itself sufficient and necessary to show that the underlying mechanism truly works; a falsified theory does not add a jot to this straightforward conclusion. Similarly, what could be the justification for rediscussing Deutsch's workable but hardly supportable theory of mental maps? (p. 350 ff.) Among the decisively falsifying facts it is sufficient to mention that some spiders, ants, honeybees, and rats are known to be capable of preplanning novel shortcuts based solely on detour experience, and hence by means of mental maps that do not match Deutsch's theory (Gorner 1958; v. Frisch 1967; Jander 1957; Tolman, Ritchie, & Kalish 1946). Nobody who is sufficiently informed will disagree with Gallistel's well-taken central point that the study of action patterns is seriously underrepresented in contemporary psychological teaching and research. One strangely human explanation for this relative neglect is the psychologist's long tradition of ignoring or even rejecting ethological theories and knowledge. As an interesting historical fact, I would like to mention that in 1954 E. v. Hoist (who figures so prominently in this book) joined with K. Lorenz, the "father of modern ethology," in founding the renowned Max-Planck Institut f iir Verhaltensphysiologie. Overall, I wish this book had been offered merely as selected readings of classic papers on the organization of action patterns. At present, a new, true psychobiological synthesis of the organization of action is nowhere visible.

A servomechanism, in turn, is defined so loosely that this supposedly elementary term correctly applies to countless definitely nonunitary behaviors such as overall feeding behavior. Finally, from among all the diverse, intrinsically driven, elementary movements only the oscillator type is considered important, without any reason being given, and all the nonrhythmic action types go unmentioned. Altogether, with all the crucial rules for setting up a set of classifying concepts being violated, I cannot imagine how anyone could succeed in understanding the organization of action patterns by employing such an unscientific terminology. If the exasperated reader tries to infer the meaning of the three elementary concepts from their use, he has to suffer further frustration. A case in point: a scratch reflex, by the definition of a reflex, should be an atomic behavioral unit. Nevertheless, this unit contains as one of its constituents a true oscillator which, by definition, is an atomic behavioral unit in itself. Adding to this muddle is the ambiguous use of "oscillator" as, on the one hand, merely a nervous mechanism generating rhythmic output and, on the other, as a unit of behavior comprising both nervous and muscular processes. Similarly, there is no way of understanding why the composite complex behavior pattern of a taxis is at one point considered an elementary behavioral unit, a servomechanism (pp. 143, 402), and at another point a more complex multiunit behavior (p. 275). If, finally, the utterly desperate reader seeks help from the glossary, his conceptual confusion will be further compounded. There the elementary unit of behavior is paraphrased as "a nerve and muscle circuit capable by itself of mediating a simple naturally occurring movement." Thus, strangely enough, reflexes and servomechanisms are now ruled out as units of action by virtue of their constituents, which are not only nerves and muscles but definitely also receptors. As the only viable solution to this plight I suggest that sensory reception, perception, and cognition be disregarded when one is building a hierarchic system of concepts to describe action patterns. Gallistel's much more ambitious attempt at developing a comprehensive system of neurophysiologically explanatory concepts for action patterns is bound to fail given our current state of pertinent ignorance. More concretely, what I have in mind is some psychobiological generalized analogue for the enviably concise and expedient system of the psycholinguist: phoneme, morpheme, word, phrase, sentence, discourse. At the theoretical (model-building) level, too, blatant violations of epistemological principles abound. Again, only a few instances can be discussed. Take the newly proposed notion of the lattice hierarchy, with its superimposed informational flow or command systems (Fig. 3.1 and Fig. 1 of Precis on p. 52). If such subsystems all act at once, as is assumed in the proposed theory, then the motor output can be nothing but a dysfunctional mishmash. As this is normally not the case in behavioral reality, this fact falsifies the theory. Long ago ethologists identified this very problem and proposed and substantiated the existence of mutual inhibition between competing subsystems as a simple, powerful, and efficient mechanism ensuring mutual exclusion between biologically incompatible sets of behavior (e.g., Iersel & Bol 1958). Ignoring this necessary and time-honored ethological theory in proposing the lattice hierarchy as an explanatory principle is a deplorable step backward in the understanding of how action is organized. Whereas there is little concern about whether the proposed lattice hierarchy works (i.e., is capable of generating an output pattern resembling real action), workability is the overriding criterion in support of the proposed Fourier model, which is supposed to explain the generation of any action pattern, no matter how complex (p. 370 ff). In the author's assessment of this model, the importance of contrary

Input-output relations in goal-directed actions

M.Jeannerod
Laboratoire do Neuropsychologie Experimentale, INSERM Unite 94, 69500 Bron, France

I felt quite in harmony with many of Gallistel's conceptions while reading his book. There is, however, one important aspect of the organization of action which I think he has overlooked. To my understanding, the motor system, even at its elementary level of organization, is a response system -i.e., a system which can hardly be described functionally without taking into account the input level. Reflexes, synergies, feedbacks, and the like are basic features of the organization of action. They represent intrinsic constraints imposed by the motor system when a goaldirected action is to be put into execution. My point is that each simple component of action should not be seen merely as an element of the large ensemble which subserves the goal, but also as a goal-directed action in itself. In other words, during the course of the execution of a goal-directed action the subject is permanently faced with subgoals to which he directs subactions. Arguments supporting this point can be drawn from the observation of simple goal-directed actions, such as when a person takes an object placed at a distance from his body. Though the resulting pattern of coordination is a single rapid movement involving all the arm segments, the action of taking an object can clearly be subdivided into at least two subactions, reaching and grasping. Reaching is distancecovering action involving proximal joints, and it entails the computation of body-centered coordinates of the object's location in space. Grasping is a fine adjustment of the hand and thefingersto the anticipated size, shape, orientation, and

628

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

Commentary / Ga\\iste\: Organization of action

use of the object, irrespective of its spatial location. This separation of a single action (taking) into two subactions (reaching and grasping) becomes possible when the input-output relationships involved in a given action are considered. If attention were focused exclusively at the output pattern generated by the motor system, the only problem raised by the study of movement would be that of finding out the principles of coordination for the different segmental muscle commands. In fact, considering the inputoutput relationships introduces the need for another functional structure in the description of action. I have used the terminology of "visuomotor channel" (Jeannerod 1981) to express the idea that different aspects of the external world (e.g., in our case, different properties of the same object) are processed by specific neural systems and transformed into specific motor commands. The visuomotor channels involved in an action are assumed to be both mutually independent and mutually dependent. Their independence has to do with the notion of parallel activation, whereby the input feeding into one channel will be ignored by the others; their dependence pertains to a necessary law of synergy between the channels to account for the coordination of the commands resulting from parallel activation. These points are illustrated by the following simple experiment. Subjects had to reach for and grasp real objects placed 40 cm away on a table in front of them. Target objects were presented through a mirror, in such a way that subjects could see only the virtual image of the object projecting at the table level. The result was a situation in which the subject reached for the virtual image below the mirror without seeing his hand and met a second, real, object placed by the experimenter at the expected location. This is usually referred to as the "visual open-loop" condition (Held & Gottlieb 1958). The particular feature of the experiment was that the target object could be suddenly modified in shape or in orientation without changing its absolute spatial location. This perturbation, which occurred randomly across trials, was produced by rotating the object at various angles with a small motor triggered by the onset of the reaching movement. By this technique, the same object (e.g., an ellipsoid object) could appear either as an ellipsoid or as a sphere. Under other conditions an elongated cylinder presented in the frontal plane could be suddenly rotated by 90 around its midpoint, thus assuming a sagittal orientation. Movements were recorded at 50 frames per second on 16-mm cinefilm. For each movement, the duration and the velocity profile of the reaching component were measured. For the grasping component the pattern of the anticipatory shaping of the fingers was also studied on the successive frames, by measuring the distance between the fingertips. The film analysis revealed that the subjects had no problem in accurately grasping objects different from those for which their movement was initially planned and directed. Furthermore, the correct adaptation of the movement to the sudden change appeared to be due to a very selective repatterning of the motor commands. The reaching component was not affected as a function of whether or not the shape of the target object was changed at the onset of the movement. The duration, maximum velocity, and shape of the velocity profile remained such that no significant difference could be detected between reaching movements executed in normal and in "perturbed" trials. This was not the case, however, for the grasping component. The anticipatory hand shaping was modified "in flight" when this was necessary for the new configuration of the object. There is apparently a complete redistribution of motor commands within the channel (or channels) affected by the perturbation at the input level. Figure 1 shows movements directed either at a spherical target object (A) or at a spherical object unexpectedly transformed into an ellipsoid object at the onset of the movement (B). On the left are represented the full-blown grasping

Figure 1.

postures corresponding to the two respective trials. On the right is an indication of the time at which a change in hand shaping can first be detected in a perturbed trial when compared to a nonperturbed trial. Although 420 msec after the perturbation has occurred the grasping posture is still similar for the two trials, a clear difference is visible some 120 msec later (540 msec). This experiment clearly shows that motor commands specified by a certain visuomotor channel can be selectively modified during the execution of the movement. In addition, the fact that a change in object shape affects only the visuomotor channel relevant for shape implies that the compensation within the perturbed channel has. to be constrained within the time limits required for execution in the other, nonperturbed channels. The temporal frame imposed on all the channels participating in a given action (as defined by the synergy rule for the movement) constrains the expression of the control of individual channels by the ongoing (visual) input. This is an example showing that the structure of motor acts directed at an external goal is not only governed by motor rules. This structure is such that it allows a continuous monitoring of the movement in terms of the goal itself. The goal shapes the parameters of the movement channel by channel at the time when the program is built up before execution and it eventually modifies these parameters during the course of execution. Input-output relationships thus determine a further organizational principle superimposed upon those described in Gallistel's book.

Dynamic servomechanisms are more fun: A critical look at chapters 6 and 7 of The organization of action
E. R. Lewis
Department of Electrical Engineering and Computer Sciences, University of California, Berkeley, Calif. 94720

With its nose ringed with infrared sensors, the Sidewinder antiaircraft missile relentlessly homes on the hot exhaust ports of its target. Constantly keeping its steering vane at an angle directly proportional to the imbalane in the inputs to its infrared sensors, it corrects for evasive maneuvers by the doomed target until the inevitable final encounter. With this vivid example, the reader is introduced to the servomechanistn, one of the three basic elements of action proposed in The organization of action. As it is described here, the Sidewinder missile belongs to a very elementary class of servomechanism, often called proportional-error servoTHE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

629

Commentary/Gallistel: Organization of action

mechanisms. Considered deterministically, such servomechanisms can home perfectly on stationary targets under ideal circumstances; but it is well known that they cannot, in principle, home perfectly on moving targets (movement in the case of the Sidewinder's target being its evasive maneuvers out of the path of the missile). Being stalked by a proportional-error Sidewinder, the wise pilot would know that the missile's corrections will lag slightly behind his own evasive actions. Therefore, he might execute a giant loop, coming out of it flying as close as possible to the ground. With any luck at all, the pursuing proportional-error Sidewinder will be underground when it finally comes out of its loop. Both the inertia of the Sidewinder and the viscosity of the air prevent the missile from turning instantaneously and tracking its moving target perfectly. Unless it can somehow anticipate the future positions of its target, it simply cannot maintain zero tracking error when that target is taking evasive action. Probably the most obvious first step in any attempt to anticipate the target's future positions is to estimate the angular velocity at which its position is diverging from the path of the missile. There are at least three clues to that velocity immediately available to the control system of the Sidewinder. One is the rate of change of the error itself (i.e., the rate of change of imbalance of inputs to the infrared detectors). Another is the rate of change of the signal driving the missile's steering vane; and the third is the rate of change of the missile's actual course. If the bending of the steering vane were in part proportional to the imbalance of inputs to the sensors and in part proportional to one of these estimates of the target's velocity of divergence from the missile's path, then the tracking error could be reduced or, under certain circumstances, even brought to zero. The point here is that with a relatively simple modification, we have made our hypothetical missile not only more effective, but also more interesting. Unless a servomechanism is operating in a quasi-static world, the dynamic factors associated with it (e.g., its own inertia and friction, the transient dynamics required of it, and the transient dynamics of the ambient elements acting upon it) are quite critical. The existing theory of servomechanisms deals extensively with such factors; and, beginning with the pioneering work of Stark and his colleagues (1959), an extensive literature has evolved dealing with these factors as they relate to neuromuscular servomechanisms. However, in The organization of action, the reader is introduced neither to these factors nor to the literature dealing with them. Instead, he is introduced to the essentially static treatments of servomechanisms by Fraenkel (1927) and by von Hoist and Mittelstaedt (1950). These treatments are elegant and definitely should be read by any student interested in neuromuscular servomechanisms; I applaud the author for including them in this book. However, they are very early endeavors and definitely do not bring the reader to the state of the art on the subject. After reading the admirably clear presentations in chapters 6 and 7, the serious student not already familiar with the study of biological servomechanisms should probe into the recent literature on the subject. It will open his mind to the possibilities of fascinating dynamic operations that might be carried out by the neural networks of the central nervous system. behavior, including the highly complex and flexible coordinations underlying speech production, I welcome Gallistel's well-written book on the organization of action. It brings together in one volume the classical problems of behavior and its causation, structure, and timing, and should prove especially useful for graduate seminars. The one classical problem which is noticeably missing in The organization of action is the problem of serial order: how do we generate preplanned sequences of behavior in proper serial order (when we do) and improper order (when we make errors). Besides neglecting Lashley's problem, Gallistel seems to have misread Lashley (1951) as providing a holisitc argument or solution to the problem of serial order, which is undermined by more recent analyses of, say, locomotion (p. 136). This neglect of Lashley seems unfortunate, not just because of the importance of the serial-order problem but because Lashley in fact championed some of the very mechanisms (even for locomotion) that Gallistel is championing now - e.g., systems of coupled oscillators, and priming (partial activation or potentiation). Lashley's (1975) emphasis on the hierarchical organization of motor mechanisms also echoes throughout Gallistel's book. However, Gallistel seems to forget about this hierarchical structure in his discussion of motor programs or schemas (p. 368-71). Like many other investigators, Gallistel assumes that the engram for a skilled movement specifies the movement trajectory in space, but not the neuromuscular activity required to achieve that movement in any one case. The reason given for this "abstract program" assumption is behavioral plasticity in phenomena such as action constancy, where a movement of the same form or function can be generated by means of grossly different patterns of muscular activity; it is, for example, possible to sign one's name with either hand, either at a desk (with muscles of the forearm and fingers) or on a blackboard (with the muscles of the shoulder). The common element generating the common form appearing in the signatures is a set of coupled oscillators or Fourier components, which are independent of "which groups of muscles are to be active, in what order, and for how long," according to Gallistel (p. 370). Recent experiments directly contradict this "abstract program" assumption: Klapp, Grein, Mendicino, & Koenig (1979) have shown that the motor-memory code for hand movements is both environmental and anatomic, as would be expected in a truly hierarchic system, where high-level components specify spatial goals and lower-level components specify muscle-movement goals. The fact that transfer from one response mechanism to another is sometimes less than perfect is also a problem for the "abstract program" assumption. Consider signing one's name with the unaccustomed hand, for example. One can transfer either the musclemovement code, resulting in a mirror-image output (see Lashley 1951), or the spatial code, resulting in an approximately normal signature. But in either case there is considerable decrement in maximal speed and dexterity of form. So, if an action sequence such as signing one's name is only represented abstractly as, say, a set of Fourier components, independent of hand or other muscle systems, what prevents perfect transfer in writing with the unaccustomed hand? And if muscle movements are not part of the long-term store, why are we able to transfer the muscle-movement code at all so as to produce mirror writing with the unaccustomed hand? The answer to the latter question is again that motor programs are hierarchic in nature, with lower-level components representing specific muscle goals and higher-level components representing spatial goals. And the answer to the question of imperfect transfer is surely that the speed and variability in the activation of motor components depends on the degree of practice. For example, the components (nodes) at the muscle-movement level for forming letters with the

Behavioral plasticity, serial order, and the motor program

Donald G. MacKay
Department of Psychology, University of California at Los Angeles, Los Angeles. Calif. 90024

As a psychologist concerned with the execution of skilled

630

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

Commentary/Gallistel: Organization of action

nonpreferred hand have not received the extensive practice of those for the preferred hand, and therefore are activated more slowly and with greater variability. By way of contrast, perfect bilateral transfer of practice occurs when the lowerlevel nodes have already received asymptotic levels of practice. An extreme example is skill at chess, which transfers readily to the accustomed hand, since the muscle-movement nodes for moving chess pieces or any other small objects have received extensive practice since early childhood. A higherlevel example involving the same principle is skill in producing a sentence at maximal rate, which transfers without decrement to producing a translation of the sentence in the other language of a fluent bilingual (see MacKay & Bowman 1969). Here the common elements represent high-level concepts, while the divergent lower-level nodes with asymptotic degrees of practice represent the words and their phonology as well as the muscle movements for realizing the speech sounds of the words. Of course, Gallistel does make reference to hierarchical organization in discussing other aspects of behavioral plasticity, such as stumble-preventing reflexes in the cat, contextdependent phototaxes of the coastal snail, and response generalization in humans. Here his account, as far as it goes, seems fundamentally correct: higher-level nodes representing concepts such as "escape shock" and "withdraw the right hand" (in Wickens's 1938 study of response generalization) must simultaneously prime several (in this case antagonistic) nodes at the muscle-movement level. These motor nodes also receive varying degrees of priming from other sources, and whatever node receives the greatest degree of priming becomes activated, triggering the corresponding muscle movement. As Gallistel points out, these "other sources" must compute that relation between body space and neuromuscular space, thereby ensuring that a signal commanding upward motion of the fingertip produces extensor excitation when the hand is oriented palm down and flexor excitation when it is oriented palm up (p. 393). What is notably absent in this account is the triggering mechanism for selecting and activating the node with the greatest degree of priming in the domain of primed motor nodes. And as Lashley (1951) pointed out, the triggering mechanism in the case of complex behaviors must have access to the syntax of the preplanned action, so that its components are activated in proper serial order. However, we must surely foregive Gallistel on this point, since neither Lashley nor anyone else has conclusively worked out the triggering mechanism for complex behaviors and since a well-written book on the remaining problems in the organization of action is surely accomplishment enough. one thing: that we will be given a theory of instrumental or operant behaviour. There is no question that we need one. As Gallistel says in the book: "When a representation of some aspect of the world has been laid down in memory, it may help to guide and control the animal's subsequent actions. The question is . . . How can knowledge control action?" (p. 335). It is a question that has haunted cognitive theorists ever since Guthrie charged Tolman with leaving his rats buried in thought. But Gallistel does not solve more than a small part of it (and that the easiest), for the solution he proposes is to explain instrumental action in terms of classical conditioning. In classical conditioning, animals learn about the relationship between a stimulus and a reinforcer, and this stimulus, by virtue of its association with the reinforcer, comes to elicit responses related to those elicited by the reinforcer itself. Although these responses were long thought to be confined to simple reflexes or autonomic reactions (such as blinking or salivating), the discovery of autoshaping (Brown & Jenkins 1968) made it clear that the movements of an animal through space could also be classically conditioned. An appetitive reinforcer may elicit approach, and an aversive reinforcer withdrawal, and both these reactions can be classically conditioned to stimuli associated with those reinforcers. Classical conditioning of approach and withdrawal may then appear to mimic true operant behaviour: that is to say, much of the behaviour observed in ostensibly operant experiments may in fact be classical in origin. When a rat is required to traverse a maze in order to obtain food in the goal box, the experimenter may believe that he has arranged an instrumental contingency between the rat's behaviour and the delivery of the reinforcer, but the contingency actually controlling the rat's behaviour may be the classical one between the stimuli of the goal box and the reinforcer. Deutsch's (1960) theory of maze learning, which Gallistel advances as a theory of instrumental action, proposes just this: various stimuli in the maze are associated with food, and each of these stimuli in turn comes to elicit an approach response. It is ingenious, plausible, and readily instantiated in mechanical terms, but it is a theory of how stimulus-reinforcer contingencies affect animals' behaviour; it does not explain how animals learn about the consequences of their actions and modify those actions accordinglyThe question is whether we need such a theory. Perhaps we could explain all cases of apparently instrumental learning in the sorts of terms proposed by Deutsch. Perhaps there is no such thing as true operant behaviour. Certainly, there have been theorists who have advanced this argument (e.g., Bindra 1978, in this journal). But Deutsch himself never did, and he was surely right. There are numerous examples of instrumental conditioning that cannot easily be reduced to the classical conditioning of approach to or withdrawal from stimuli associated with appetitive or aversive reinforcers. Both Miller & Konorski (1928) and Grindley (1932) saw clearly the sort of task that was required to establish that instrumental conditioning was a process distinct from classical conditioning, and both studies showed that a restrained dog or guinea pig could learn to perform an arbitrary, discrete response, such as lifting a leg or turning the head, in order to obtain food. There are other, more modern examples (cf. Mackintosh & Dickinson, 1979). A plausible account of such cases of genuine instrumental conditioning, and one of which, from brief hints, one suspects Gallistel would approve, is that animals can learn that their actions are associated with certain consequences, just as, in the case of classical conditioning, they learn that external stimuli are associated with particular consequences. But the problem then becomes one of explaining how this knowledge affects action. How does a representation of the relationship between an action and its consequences determine the performance of that action?

Where's the action?

N. J. Mackintosh
Laboratory of Experimental Psychology, University ot Sussex, Brighton, BN1 9QG, England

Gallistel has brought together an important group of papers and has written a fine book about them. He is right to claim that classical physiology and neuroethology have developed a number of principles underlying the organization of patterns of movement, and these principles undoubtedly help to resolve some of the difficulties involved in developing a comprehensive theory of action - but not, I suspect, all. I shall confine my remarks to a single issue. The book deals, we are told on the jacket, with "the principles by which complex, purposive, and intelligent behavior is generated." To a student of animal learning or conditioning this can imply but

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

631

Commentary/Gallistel:

Organization of action the train that we discount all these cues. An equally compelling illusion is created by 360 movies taken from a moving vehicle. Our experience tells us that trees and houses do not stream past; it must be ourselves who move. Clearly, at the human level, past experience is more important than any efference copy in determining how we interpret visual input. We are always aware of the true orientation of visual objects in relation to our body. The lack of perceived motion does not fool us into thinking that what was straight ahead before an eye movement is still straight ahead afterwards (though it is not obvious that von Hoist realised this). It seems that there must be a side branch of the visual system devoted to the detection of movement and the efference copy must translate the whole visual image in that branch to the position it is expected to assume after the movement. If it does not appear in that position there is a perception of movement and an alarm is raised to draw attention to the phenomenon. Translation of the visual field requires a more complicated computational network than simple subtraction of motor commands. My guess is that it involves the superior colliculus, the pons, and probably the cerebellum, which is richly supplied with positional information, both efferent and reafferent. The efference-copy idea is an important one, I am sorry that Gallistel chose to present von Hoist's very confusing treatment of it. Finally, I would like to comment briefly on Gallistel's speculations on how cognitive maps are read. It is a problem that I have spent many years contemplating. Useful as the idea of the cognitive map has proved to be, I suspect that Tolman may have done psychological theory a disservice when he coined the expression. It is too tempting to take the metaphor literally and imagine something like a road map spread out on the hippocampus, or wherever. From there it is a short step to imagining a map reader who knows where he wants to go and consults the map to find out what to do. In geographical maps response instructions are implicit, or coded, but there is no reason that the brain should use the same code. It probably contains much more explicit information. In fact, I believe the "map" consists of a large number of statements or response associations of the form "Go 2 km along this route and you will come to Howard Johnson's" or "Bark twice and the door will open." A better name for this might be a gazetteer of expectancies. I do not believe this map or gazetteer is consulted to discover the response; rather, it is the other way round. Every situation presents a choice of responses, and the gazetteer is consulted to discover the consequences of each. Because the animal already has a plan for a response, the question of how to translate from a map to a response does not arise. If the associations with the planned (or potentiated) response are found to be of no value for reaching the goal, that plan is not executed, and the gazetteer is once more consulted to discover where another of the possible responses will lead. When a response is found whose outcome has associations with the goal, the potentiated response is given the green light and activated. If any of the possible responses have no associations, the animal may explore so that some are acquired, and these associations are then entered into the gazetteer for future reference. Of course, animals cannot go through this vicarious trialand-error rigmarole every time they make a response. In familiar situations the response first thought of is the one most often made before, so it is tested first and may even be released automatically, without any consultation. If that sounds like our old bete noire, stimulus-response association, that is just what it is. The fact that it is no good for explaining

A small fly in some beneficial ointment

P. M. Milner
Department of Psychology. McGill University, Montreal. P.O.. Canada H3A 1B1

Gallistel has provided an excellent synthesis of all the points that I consider most important for understanding behavior. I could document my concurrence with his conclusions by referring to several of my papers. For example, in 1959 I gave a talk that described just such a hierarchical model as Gallistel presents, with preorganized action systems for walking, grasping, chewing, and so forth, and illustrated it with a diagram of a six-tier response mechanism with a drive at the top level and motor units at the bottom (Milner, 1961; Fig. 4, p. 120). Each level facilitated (i.e., potentiated) several alternative action units at the next lower level, one of which was then selected and fired by sensory input available at that level, and so on down to the motor units where the sensory input was mainly from muscle spindles (less controversial then than they are today). The talk was published in a book somewhat ironically called Current trends in psychological theory. Obviously the trend was not as current twenty years ago as I imagined. Perhaps it was to avoid having his book dismissed as too speculative that Gallistel chose to include the long (and sometimes long-winded) classical articles. His own style is clear and lively, and the book would have been more readable if he had written it all himself. In fact, the only fly of any significance that I found in the ointment was introduced by two of his eminent coauthors, von Hoist and Mittelstaedt (1950). Von Hoist is puzzled that flies are not paralyzed by the optokinetic reflex and postulates that the reflex is cancelled by a copy of the motor command for self-produced movements. But the optokinetic reflex is relatively slow; a fly can almost certainly move too fast for the reflex to follow. Moreover, if the visual world moves suddenly, flies do not sit around readjusting their posture; they depart from the scene with as much brio as they can muster. The question von Hoist should have asked is whyfliesdo not scare themselves into fits every time they move. No self-respecting ethologist would pay any heed to what a fly did in the optokinetic-reflex apparatus. In real life the movement of a visual image depends on the orientation and distance of the objects in the field, as well as on the velocity of flight. How can a simple efference copy compensate for such multiply determined and variable input? Finally, the experiment cited to dispose of the alternative theory that the reflex is depotentiated during movement is quite inconclusive. The fly with inverted head is gyrating under the influence of the optokinetic reflex (malfunctioning though it may be); it is not making "spontaneous" movements. Neither theory predicts interference with reflexes generated by outside movement, so the fly's predicament does not help us to decide between them. At the human level it is false to assume that only the efference copy allows us to decide whether a movement is objective or self-produced. Most regular railway travellers have experienced the illusion produced by the departure of a train on the adjacent line. The feeling of movement as one looks out of the window is quite compelling and there is an almost physical shock as the last carriage passes and the stationary scene behind is revealed, bringing us to a sudden halt. We succumb to this illusion in spite of the lack of efference or any nonvisual sensory indication (vestibular, pressure from the seat, etc.) that we are moving. We know also that the train outside is not a fixture, yet we are so used to interpreting movement outside the window as movement of

632

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

Commentary/Gallistel: Organization of action

latent learning or problem solving does not mean that S-R learning never occurs under any circumstances. It is as good a model as any for explaining automatic behavior. Deutsch's (1960) theory of problem solving, which is the one presented by Gallistel, has several limitations, the most serious being that it sidesteps the general question of response translation by assuming a single, all-purpose approach response, elicited when a stimulus whose map reference receives drive activity appears. If something other than approach is required (tail wagging, for example), there is no way of representing that on the map, and no way of reading it off the map. Deutsch's model is useful for illustrating the difficulties that must be tackled and the general principles to be considered in a solution, but it is not easy to see how the concepts translate into physiological mechanisms. It is probably clear from my commentary that I do not think Gallistel has quite written the perfect book on the subject of action systems, but it is close. The main ideas are all there, and I am delighted to see them set forth so articulately. I hope the publication will ensure that this way of thinking about psychological problems will be more widely accepted in the future. Sherrington and von Hoist. Through the course of nine chapters, a series of experimental observations and mechanical analogies is used to show how three very basic organizational structures, the reflex, the oscillator, and the servomechanism, interact to generate behaviors more complex than produced by each separately. Perhaps the finest achievement of the book is in building these ideas into several principles of higher-level interaction, principles which are often alluded to elsewhere in the literature but with less than full understanding. Specifically, Gallistel describes how "competitive" interactions among structural units operating at the same hierarchical level, from whence arise the "lattices" characteristic of the hierarchy, yields a much wider range of possible behaviors. Furthermore, he demonstrates how selective higher-level potentiation of lower-level behaviors, as compared to direct command of these behaviors, enables a system that is fixed structurally nevertheless to exhibit highly flexible and perhaps "intelligent" behaviors when it operates within a changing environment. In summary, The organization of action: A new synthesis will be an excellent conceptual source book for the student interested in organized actions of the central nervous system. It should also prove of value to the researcher who, over the years, has had piecemeal exposure to concepts of organization, but has not had the opportunity to "solidify" these concepts within his own mind.

Hierarchical structures in the organization of motor behaviors

Lewis M. Nashner
Neurological Sciences Institute. Good Samaritan Hospital & Medical Center. Portland, Ore. 97209

A basis for action

Allen Newell
Computer Science Department, Carnegie-Mellon University, Pittsburgh, Pa. 15213

The conceptual emphasis of Gallistel's book elicits my commentary, not so much upon the myriad of important but specialized details as upon the ability of the book to accomplish one of the expressed purposes of the author: "to convince psychologists and behavioral neurobiologists that the conceptual gap between muscles and motivation is not the yawning chasm most of us imagine." Gallistel has assumed the role of advocate rather than reviewer or critic. He has carefully chosen a series of historically important papers based upon motor-system research and has interwoven these with his own thoughts to advance the concept of a "lattice hierarchy." At a time when our information concerning anatomical components seems to outpace our ability to comprehend the workings of the system, in-depth presentation of a physiologically relevant organization concept is extremely useful. Because this book represents the effort of one mind rather than a collection of more loosely connected chapters by separate individuals, its cohesiveness more than compensates for a lack of representation of alternative viewpoints. For example, Gallistel quietly but in no uncertain terms dismisses the reductionist approach to the understanding of CNS organization. The reader will find a natural subdivision of the book into two parts, although this separation is not reflected in its formal organization. In the opinion of this reviewer, the first subdivision is most successful in accomplishing the stated objectives of the author; it carefully builds upon very basic principles of motor organization in order to demonstrate the conceptual power of hierarchical principles. The second subdivision is smaller in scope and less exacting in its attempt to generalize concepts derived from motor organization to encompass cognitive levels of action such as motivation and learning. While this section will probably not convince the reader to the same extent as the previous one, it is useful in providing a stimulus for further thought. The first subdivision of the book is a sequence of carefully orchestrated articles by such well-known psychologists as

Gallistel puts forth his book as providing the basis on which we scientists should organize our future actions. It is crafted to capture our attention, communicating by its form as well as by its content, and containing enough contrasts to confound a simple reading. I found the result pleasing, by and large. The book strongly invites attention to its overall message, rather than to its scientific details. On its own account, by many caveats, though the scientific content is to be taken seriously, it is fundamentally in the service of the overall message. I read the message as follows: (1) The central ingredients exist for constructing a theory of the organized actions of animals, including man. (2) These ingreidents form a basis, in the sense that from them can be composed all the actions of the animal. (3) These ingredients have been available in the scientific literature for a long time. (4) Yet this synthesis is new, because we cognitive scientists have failed to recognize that these ideas exist. I found pleasing the use of original articles to convey the main substantive thrust of these ingredients. It is good to be reminded, with the pungency of the originals, of relevant work one already knew (for me, Sherrington on reflexes and Weiss on salamander transplants). It is doubly good to be introduced to neat work from the past that had somehow slipped by (von Hoist on fish fins) or that had lain beyond one's earlier scientific horizon (Wilson on insect gaits). It is even good to have one's position challenged on work long since studied and set aside (the Deutsch model), even if old opinions more or less prevail. The substantive proposition of the book can be captured almost entirely by enumerating the proposed basis. Its elements are the reflex, the oscillator, and the servo. Its main organizing principle is the control hierarchy. Its control primitive is potentiation and depotentiation. There is also a

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

633

Commentary/Gallistel: Organization of action

principle of the clever simple solution (illustrated most strongly by insect locomotion): simple mechanisms can solve amazingly complex behavioral problems for the organism. This principle bolsters the view that a simple (bio)mechano set suffices for obtaining intelligent action in the face of a variable environment. At this point I do not know quite what to say. I'm afraid I am about to carp. Since I feel almost entirely positive about the book and its potential effects, this is definitely not a good thing. Nevertheless, here is the carp - and then I'll try to rise above it. Gallistel does not practice what he preaches. At the level of the message, everything that he says is familiar in cognitive psychology. Not only that, but it has been clear for twenty years. There is no problem with the homunculus; with the composition of action out of a procedural basis; with whether mechanisms can yield intelligent and purposive action; with autonomous signal generators, hierarchy, multiply evoked subsystems, gated control, generalized control systems, mechanically realizable representations; and so on. For all of this, there is no problem at the level at which he treats the issues. For someone who casts his book to chide the field for not heeding existing science - as in his contrapuntal treatment of old writings and declared new synthesis he has not himself attended closely to what is known and understood in the field he seeks to instruct. I have stressed the level of the message. At levels of specificity, detail, and task scope beyond those in the book, there are important issues on all the above topics. The difficulty is not a failure to be concrete; the book succeeds there admirably. The difficulty is that it is illustrative. In fact, the trouble is that the book is a handwave - well executed and based on real data, but a handwave nonetheless. It is reminiscent of Hebb's Organization of behavior (1949), a work remarkably similar in motive, style, and even argument trying to get psychologists to take the neural level seriously. (It is not cited, interestingly enough.) Hebb's was one of the great handwaves in psychological history, and its impact was immense and deserved. At that historical moment, Hebb's overall message was a breath of fresh air. The new orientation counted; the details didn't. In a kiloword review I cannot document my claim that the cognitive world already understands the macro lessons that Gallistel seeks to tell us. Also, I am not my brother cognitive psychologist's keeper. I will not aver that all psychologists know what they should about modern artificial intelligence, robotics, computer science, and control engineering, to inform their more direct knowledge of cognitive models and data. But there are plenty that do. What would I have liked? Perhaps a detailed critique of the biological plausibility of Saltzman's (1979) attempt at synthesis of sensorimotor representations (though that may have emerged too late). Perhaps a discussion of the nature of potentiation as one type of control primitive against what we know of the total needs for control, say, from programming languages. Perhaps a discussion of the biological plausibility of various bases for composing functions, out of the indefinite variety of ways that are known for this besides Fourier decomposition (e.g., all the orthonomal polynomials, many other transforms, etc.). Perhaps even my old favorite (Newell 1980) of how the type of action system under discussion interfaces or melds into a symbolic system. Mostly, I think, we are ready for actual theory that can be seriously taken to task and then improved. We do not need to be convinced at the level at which Gallistel seeks to convince us. But enough. The motor system certainly has a distributed continuous control. It is due for much more attention in the coming decade. And much will be learned thereby that will affect and enrich our view of the total cognitive system. Gallistel makes the case that the time is ripe (to squeeze an old cliche). And his work will certainly help bring on this next wave of research.

Behavioral flexibility and the organization of action

David S. Olton
Department of Psychology, The Johns Hopkins University, Baltimore, Md. 21218

Gallistel's "new synthesis" is a welcome addition to the currently available descriptions of animal behavior. Although many of the earlier models can be adapated in some post hoc way to account for the types of behavior discussed by Gallistel, these models still have difficulties because they were never designed to predict the flexible, organized aspects of behavior which are the focus of the work here. As examples of these intelligent behaviors become more and more frequent, both from the laboratory and from the animals' natural habitats (see, for example, Kamil & Sargent 1981), we desire a description of behavior that focuses directly on them, rather than including them as an afterthought. I was particularly interested in the discussion in chapter 11. The idea that knowledge (nodes and connections) in a cognitive map can exist without specifying any action (direction of movement) is an excellent one, particularly in the context of the more cognitive emphasis that currently characterizes psychology. The question, of course, is how we get from these unobservable cognitive processes to observable behavior. Gallistel says: "If we know both what an organism wants and what it knows, then we feel intuitively that we can predict what the organism will do" (p. 351). But how well do these intuitive feelings actually lead us to the correct predictions? In some cases, documented in chapter 11, appropriate knowledge plus motivation does produce intelligent action, such as taking shortcuts to the desired goal (see also Becker & Olton, in press; Becker, Olton, Anderson & Breitinger, in press). But such is not always the case, and the literature is replete with examples of animals exhibiting very inflexible and unintelligent types of behavior. Some of these examples may be found after specific training, as in Maier's "frustration" experiments with the Lashley jumping stand (Maier 1949). For some rats, the first discrimination was solvable (i.e., one stimulus was always correct). These rats rapidly learned to jump only to the correct stimulus. For other rats, the first discrimination was not solvable (i.e., each stimulus was sometimes correct and sometimes incorrect, and the rat was unable to predict which would be the case for any given trial). When subsequently given the discrimination that could be solved, these rats persisted in choosing incorrectly, in spite of indications (such as changes in latency and posture when making choices) that they had learned which stimulus was correct and which was incorrect. Inappropriate choice behavior has been documented in a variety of other situations as well (see Olton 1979, pp. 584-85, for examples). Most importantly, these experiments did not include any specific training experience (like that of Maier's [1949] experiments) prior to the discrimination task. Nonetheless, the animals were notably inflexible, running into walls or off mazes, much as the more stereotyped, mechanical models of animal behavior might have predicted. Finally, a particular animal can be trained to exhibit consistently either a relatively flexible or a relatively rigid behavior in a variety of different experimental situations. The ideas of learning set, transfer of training, functional fixedness, learned helplessness, learned success, etc. all imply a general propensity toward either insightful, adaptive, and intelligent

634

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

Commentary/GaMisteh Organization of action

new approaches to a problem or a rigid, maladaptive, and inappropriate perservation of old habits. Thus I ask the following question: when knowledge and motivation combine to produce action, what are the variables determining how adaptive and flexible this action will be? Something other than the components of the present theory seems to be required, something that reflects the "creativity" with which an animal approaches a given situation. How would Callistel incorporate such a variable into his organization of action? How is it established, and does it affect the direction of action when behavior ensues? depend upon environmental "instruction" as a significant determinant in the genesis of neuronal circuitry. At a more molar level, movement is also important in determining the environment and experience of a given organism. Piaget realized this and continually stressed the interdependence of motor and cognitive development. This interplay was central to his view of the child as experimenter, not passive metaphysician. Gallistel's book should help bring motor processes and related topics to the attention of psychologists and make relevant areas of psychology intelligible to neurophysiologists. The book makes contributions at several levels. It provides thoughtfully selected papers and excerpts from the history of motor control (e.g., Sherrington, von Hoist, Weiss, Wilson, Fraenkel), some of which are relatively unknown or difficult to obtain. Gallistel's perceptive introductions and analyses make the reprinted material especially useful. He successfully translates early ideas and terms into the parlance of contemporary neuroscience. An excellent glossary of frequently used but often undefined terms is even included. A second contribution is the idea that behavioral control is exercised through the selective potentiation of functionally coherent subsets of lower units of action. This approach provides a common ground for psychologists and those with more molecular interests in behavioral neuroscience. Other topics covered are central to motivational theory, perception, cognitive psychology, development, and ethology. I highly recommend Gallistel's book as a unique and useful historical introduction, source book, and synthesis of a wide range of topics concerning "the organization of action." It deserves a wide audience.

Giving behavior to psychology

Robert R. Provine
Department of Psychology, University of Maryland Baltimore County, Calonsville, Md. 21228

In response to the behaviorist revolution, it has been said that "psychology, having first bargained away its soul and then gone out of its mind, seems now . . . to have lost all consciousness" (Burt 1962, p. 229). Amid the turmoil about psychology's state of mind, it went unnoticed that there was little "behavior" in behaviorism or elsewhere in psychology. With few exceptions (e.g., linguistics and species-typical behavior), contemporary psychology is devoid of fine-grain analyses of motor behavior. Where is our choreography or grammar of movement? Where is our "abnormal psychology" of movement disorders? The motor chapter of the typical textbook of physiological psychology has little to offer but a tired description of the alpha and gamma motor systems and simple and complex reflexes. These cursory treatments, apparently added only for "completeness," bear no relevance to the rest of the book. They do not even mention the contemporary research on endogenous pattern generators that has revolutionized thinking about motor control. Man and beast are instead treated merely as reflex machines. The empirical philosophy which gave rise to modern Western psychology produced a vigorous psychophysics and psychology of perception, learning, and cognition. Sadly, on the motor side, the empirical tradition generated only that meager parcel of motor behavior, the response. The relative neglect of motor processes also stems from the fact that the development of movement and its neuromuscular "machinery" is relatively independent of experience, (Provine 1979, in press) and that once movement disorders are produced the prognosis for recovery is often poor. Motor behavior has not earned a significant place in a psychology devoted to cataloguing environmental influences on behavior. Motor processes deserve a more prominent position in the behavioral sciences. Indeed, the senses and the nervous system evolved in the service of movement; without movement, they would confer no selection advantage. There are no intelligent turnips. Neuroembryological analyses also indicate a unique and important position for the motor system (Provine 1976). The precocity of motor processes is unrecognized by most developmental psychologists, although it was first noted by Preyer (1885) almost a century ago. Embryos often act before they react ("spond" before they respond), and spontaneous (nonevoked) processes are probably responsible for most embryonic motility. Furthermore, the efferents of some central neuronal circuits develop before they receive their afferents (Bruce & Tatton 1980; Stein, Clamann, & Goldberg 1980). The sensory and supraspinal systems may not "capture" and modulate the freewheeling motor systems until the basic structure of the nervous system is well blocked out. These behavioral, physiological, and anatomical findings force a reconsideration of developmental accounts that

Can mental representations cause behavior?

Edward S. Reed
Center for Research in Human Learning, University of Minnesota, Minneapolis, Minn. 55455

Gallistel's book is an extremely important contribution to the theory of action. He has succeeded admirably in synthesizing several centuries' worth of knowledge into a clear presentation. Although my comments below will be highly critical, they should not be seen as detracting from the great value of Gallistel's work. Without Gallistel's lucid and comprehensive review, the important issues I attempt to discuss below could never have been raised. Gallistel's work is an impressive synthesis, but it is by no means "new," as he claims; he is misled by the widespread mythology of the "novelty" of psychobiological ideas which are at least three centuries old (Reed 1980). This misapprehension obscures deep problems in Gallistel's theory. Gallistel has set out to answer, in as scientifically materialist a way as possible (p. 4 ff.), the question Descartes first raised in his Passions of the soul (1649): how do ideas (mental representations) cause behavior? Like Descartes, Gallistel proposes a two-system answer: the cognitive system selectively potentiates the machinery of the body. We are told (p. 358) that, although no one has ever had the faintest idea how ideas could cause actions, this is of no great concern to the theory. Personally, I think three centuries of inability to answer a scientific question demonstrate that the question itself is suspect. Certainly, Gallistel should address this perennial failure in a more serious manner than his vague and desultory exposition on pages 388-90. But even granting that Gallistel may yet solve the problem that defeated Descartes, Kant, Miiller, Sherrington, and others, there are further problems with this old synthesis in new guise.

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

635

Commentary/Gallistel: Organization of action

The most frequent reason given for hypothesizing that mental representations cause behavior is an argument from exclusion (see Darwin 1958, p. 84). For example, Gallistel claims that because no other explanations of Maier's (1929) experiments are viable, cognitive maps must be hypothesized to explain rat detour behavior (p. 338). Crudely put, the generalized form of this argument is: mental representations are hypothesized to account for whatever knowledge is possessed by animals that is not explicable by appeals to habit and experience. Representations are nowadays invariably supposed to explain novel and insightful behavior, instances of action in which animals in no way repeat earlier episodes of behavior. Like the predictions of a scientific theory, representations are conditionals ("If I do X, then Y will occur") and can support counterfactuals (X need not be the case; Fodor 1975). These aspects of representations thus provide an appearance of explanatory power. For instance, the hypothesis that an animal possesses a cognitive map seemingly accounts for all the spatial knowledge that animal has, especially its knowledge of the areas of the environment not now being experienced ("If I move in that direction, then I will experience something like X"). There are at least two serious difficulties with this sort of exclusion argument for representations. First, it makes cognitive psychology parasitic on behaviorism. Ever since Chomsky (1959), cognitivists have been hypothesizing mental representations whenever their favorite (or most reviled) behaviorist's theory fails. Chomsky's arguments against behaviorism are largely correct, but Chomskyan accounts of language learning have been seriously weakened by this absurd strategy of standing Skinner on his head. If behaviorism is wrong, then behaviorist accounts of experience are wrong and nothing can be inferred from failures to explain using those accounts. Hypothesizing entities, such as mental representations, is a perfectly legitimate research strategy, but one's hypotheses should be designed to generate tests and to explain facts, not to explain away previous ignorance. [See also Chomsky: "Rules and Representations" BBS 3(1) 1980.] The second problem with basing arguments for mental representations on the poverty of experience is much more serious than the first one. By its very structure, the argument from exclusion makes the origin of representations (in ontogeny or phylogeny) supernatural. It is not possible for mental representations with the strength of counterfactual conditionals to have evolved, as a simple example, drawn from Lee (1980), will show. Cannets are seabirds who prey on fish by diving to catch fish swimming near the water's surface. In diving, what gannets usually do is drop, allowing gravity to accelerate their bodies while they use their wings to steer. In order to avoid deadly damage to their bodies, these birds must retract their wings before the wing-trunk junction hits water. However, if they retract their wings even a brief moment too soon prior to entering the water, they will miss their prey. In short, for gannets, there is an extremely brief "window" of success in fishing. How can this single, but precise, behavior of wing retraction be explained? Gallistel apparently would explain gannets' wing-retraction ability by hypothesizing that these birds have a spatial or temporal representation of their diving trajectory, a representation which selectively potentiates reflex mechanisms of wing closure. For each dive the birds have to compute a representation with the following content (but not form): "If I continue to fall at the present rate I will hit the water at time t, so I must retract my wings at t-A." How could these birds have evolved the ability to represent this counterfactual? Birds who actually experienced hitting the water would have died before they could live to tell their tales. This selective mortality can explain why gannets rarely smash themselves to death, but it cannot explain how they could use representations predicting that danger to control their actions. Although the hypothesis that animals form mental representations or "models" of the world (Pantin 1965) is widely adhered to, no one has ever explained how those "models" could convey information about nonactual states of affairs ("If I don't hide, that tiger will eat me, so I hide."). Yet, according to Gallistel, representations are to be hypothesized precisely in order to explain how animals come to know not-yet-actual or not-yet-experienced states of affairs. To be sure, some representations simply store past experiences (memory), but the explanatory power of representational hypotheses comes from the alleged ability of representations to go beyond the information given, and therefore to explain novel behavior in animals. For example, Gallistel (p. 358) and others wonder whether the rat's representation of space is "Euclidean." What rat (or person) has ever experienced two parallel lines extended to infinity and then checked to see that they nowhere intersect? How might a rat (or a person) have invented such a hypothesis in the first place? Cognitivists, like Gallistel, agree with behaviorists that hypotheses must come from experience or from heredity. Ex hypothesi, representations with the complexity of geometry do not come from experience; hence, they are (at least in part) innate. But this just makes matters worse: for natural selection to have produced rats who hold to Euclidean geometry there must have been, at some time, a population of rats whose representations of space varied, so that selection could operate and pick out the most successful representation. So Gallistel must now explain the origin of rats whose representation of space is Euclidean, Lobachevskian, Riemannian, and so on! I take it that the above is a reductio ad absurdum, not of mental-representation hypotheses as such, but of the argument that representations (or animals) go beyond the information given. Animals act as if they were considering counterfactuals ("If I do X, then Y; not X, then Z; Y is preferable to Z, so I will X"), but this does not entail that animals go beyond their experience, or that they have mental representations. Rather, it implies that perceptual information is far richer than behaviorists or cognitivists have allowed, and that perception itself supports these counterfactual seeming episodes. Gannets, to return to Lee's (1980) account, control their wing retractions by initiating movement according to a margin value of a dimensionless optical invariant r'(t) specifying whether their dive will be successful. Any system whose behavior is a function of r(t) or r'(t) (I call it the "Lee number") acts as if it were predicting a future possible state of affairs. Whether or not gannets represent Lee numbers to themselves is anybody's guess, and irrelevant here, because, representation or not, the explanation is due to the Lee number. What is relevant is that, by changing their behavior when -r(t) reaches a marginal value, the birds change their actual flight course, so that the previous Lee number specified a potential, not an actual, collision. (For a more detailed discussion, see Turvey, Shaw, Reed, & Mace, in press.) Note also that Lee numbers are an environmental resource (being facts of opticalflow)to which animals might become adapted by natural selection. The comparative zoology of eyes in fact suggests the existence of strong selection pressures for the development of visual systems capable of detecting optical flow (see Land 1980; Salvine-Plawen & Mayr 1977). From the biological and materialist point of view espoused here, and apparently favored by Gallistel, the strategy of hypothesizing representations that go beyond the information given is pernicious: it can only explain why animals live in fantasy worlds, not the real, evolving world. If Gallistel ever solves Descartes's question about how representations guide actions, the representations involved will be accurately encoded models of the real environment, and the information within them that enables animals to anticipate the consequences of their behavior.

636

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

Commentary/Gallistel: Organization of action Behavior ignored

Peter C. Reynolds
1660 North LmSalla Slroel, Chicago, III. 60614

Although Gallistel's book is less a synthesis than a compendium, it does cover old ground in a competent and clearly explicated way: hierarchy and heterarchy, reflex and servomechanism, coupled .oscillators, and Fourier transformations are all here, as are those oft-reprinted authors, Weiss, von Hoist, and Sherrington, Yet, far from going beyond these traditional concepts, Gallistel stays firmly within the rationalist, individualist, and ahistorical framework of American experimental psychology. Gallistel's animals have no social behavior, and they live in an ecological vacuum. They have motivation but no emotion. They have motor skill but no communication. Ethological authors are cited to bolster the critique of naive behaviorism and reflexology, but the main import of the ethological approach - that behavior has a social context and a natural history - is nowhere to be found. The environment intrudes in the case of the digger wasp, but only to illustrate the abstract concept of mental maps; and hippocampal function - the only excursion into limbic forebrain mechanisms - is reduced to the point-by-point representation of Euclidean space. There is not even a hint that social action cannot be reduced to the potentiation of motor acts within the single brain, much less awareness that social action has organizational principles of its own. Moreover, Gallistel's view of cognition is unclouded by a single ethological doubt: phyletic history and ecological niche may constrain the motor patterns that can be operantly conditioned, but they never raise the possibility that the laws of thought might be species-specific. These oversights are not inconsequential to a theory that sets out to bridge the gap between motivation and behavior and to explicate the organization of action. They reflect the failure to come genuinely to terms with the ethological critique of traditional experimental psychology. Ethologists experiment, but they experiment with behavior whose naturalistic context is known. They do not assume that theories of behavior are to be solely derived from isolated physiological preparations confined to the laboratory bench. Gallistel's book, by failing to relate the physiological concepts to the larger context of animal behavior, only tilts with ethological ideas while consistently skirting the real methodological issues.
ACKNOWLEDGMENT This work is produced by aid of a grant from the Harry Frank Guggenheim Foundation.

Gallistel's metatheory of action

H. L. Roitblat
Department of Psychology, Columbia University, New York, N.Y. 10027

Gallistel's presentation of a "new synthesis" represents an important contribution to our understanding of the factors that control behavior. It represents not so much a theory of action as a framework in which to cast such a theory. It is a "new way of seeing" (Kuhn 1962) behavioral phenomena resulting from a modification of behaviorism. One of the fundamental assumptions of behaviorism is that of a direct correspondence between the overt features of behavior change - antecedent stimuli and resulting responses - and the covert internal changes responsible for learning. As Gallistel makes clear, this assumption allows only a single level of organization, a direct connection between

stimuli and responses (S-R associations), itself having no internal structure. This single-level control structure means that all behavior is controlled by reflexlike associations and that all responses are elicited by external stimuli. Gallistel replaces this assumption of external correspondence with a similar assumption of internal correspondence involving at least three types of fundamental units - reflexes, oscillators, and servomechanisms - each under the control of other elements arranged in a lattice hierarchy. In so doing, he has solved a number of the most nagging problems confronting behaviorism. For example, the inclusion of endogenous oscillators removes any problem generated by the apparent spontaneity of behavior when external eliciting stimuli cannot be identified. The relationship between motivation and performance is solved through the inclusion of potentiation and depotentiation mechanisms that govern both motivation and performance. The problem of level of analysis in behavior - that is, the controversy over whether animals learn muscle twitches or acts (operants) - is replaced by the hierarchy principle according to which learning can take place at practically any level. Despite these innovations, Gallistel's synthesis is still only a modification of behaviorism. His strong commitment to reductionism and physical realizability permits nothing else. All of the principles and mechanisms that control behavior reduce to interconnections among three fundamental units, sometimes augmented by apparently ad hoc modifications such as the autonomous buildup of specific potentiation. While this reductionism has shown itself to be quite powerful in dealing with the organization of action at some levels, additional principles are undoubtedly necessary to deal with the level of behavioral adaptation usually studies in the psychological laboratory. For example, the only mechanisms explicitly in the hierarchy that could account for learning are the selective potentiation of reflexes and servomechanisms and the selective coupling of oscillators. Alone, these mechanisms do not constitute explanations. The task of any learning theory, whatever its scope, is to explain the particular changes that occur. The hierarchy principle and its associated mechanisms are not sufficient to explain even the results of Wickens's (1939) finger-flexion experiment. Different subjects adapted to the experimental demands at different levels of the hierarchy, one learning a particular reflex, the others learning an action integrated at a higher level. The failure of Gallistel's synthesis as a theory of learning should not be taken as a strong criticism of the system. It is not intended as a theory of learning, and in fact it is not even a theory of action. Examples of the hierarchy principle can be found ^everywhere. Because many actions involve the same sense organs and the same muscles with their innervating neurons, the principle of a lattice hierarchy must be valid at some level. Furthermore, practically any theory of action can be described in terms compatible with a lattice hierarchy. Precisely this ubiquity prevents it from having explanatory power. A second problem (perhaps "lacuna" is a better term) is in Gallistel's treatment of representations. It seems clear that much of behavior is not easily describable in simple stimulusresponse terms but seems to require the use of more elaborate representations of knowledge. The question, then, is how the system that represents knowledge affects the system that controls action. Gallistel seems to attempt a solution to the problem by proposing two separate systems: a cognitive system, representing knowledge, and an action system. On closer examination, however, it becomes clear that these are really different levels in the same system. That is, Gallistel adopts the same solution to the homunculus problem adopted by the behaviorists, that of including knowledge as a part of the action

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

637

Commentary/GaWisteh Organization of action

system. His solution is a modification of the behaviorist position in that knowledge is assumed to be more complex than simple changes in response propensities. The prediction of the behavior of an animal requires the specification of its knowledge and its wants or goals. Callistel assumes that motivation spreads from goal-related motivational units high in the hierarchy through the map to selectively potentiate action units at the level of telotaxes. Representations, like other units of action, consist of interconnected nodes interacting with lower levels. While unification of knowledge and action systems, like the unification of motivation and action systems, has distinct advantages, such as leading to articulatory coding schemes, the available evidence appears to demand more complex representational systems containing dimensions orthogonal to the action system (Roitblat 1980; 1981). Hulse & Dorsky (1979), for example, present quite strong evidence that rats extract and represent information about a sequence of reinforcer values that reflects the overall rule structure of the pattern. Learning is faster with monotonic than with nonmonotonic patterns, and there is a positive transfer from one pattern to another employing the same rule. Despite arguments by Capaldi and his associates (Capaldi, Verry, & Davidson 1980) to the contrary, it is unlikely that this information could be stored in anything resembling an action hierarchy. This point is further illustrated by the work on the rotation of mental images done by Shepard and his associates (Shepard 1975; Cooper & Shepard 1979). Callistel recognizes that the performance obtained with this task requires an elaboration of the hierarchical representational scheme to include dimensionally encoded vectors. Even this elaboration, however, is not sufficient to account for the finding that greater degrees of "mental rotation" take more time without the ad hoc assumption that values in the dimensional vector can only be adjusted in a stepwise fashion. These experiments appear to require a representational system that is orthogonal to the action system, not one "flowing" directly into it. In summary, Callistel does present a new synthesis of far-reaching importance. Whatever difficulties are present appear primarily to be the result of attempts to make the system operate as a theory of behavior. It is not a theory but more of a metatheory, suggesting how theories of action might be organized and providing a framework in which to cast theories of behavior. It will certainly require elaboration, including more kinds of higher-level structures and the specification of the particular laws that control the hierarchical interaction among units. ACKNOWLEDGMENT Preparation of this manuscript was supported by NSF grant BNS 7914212. did no justice to the expository powers of the original writers, or to the background of their contributions. Most important, these different conceptions of the functional units of behavior were difficult to integrate into a common scheme. They were largely based on unconnected pieces of research using (for psychologists) peculiar subjects and procedures, and the results were considered primarily relevant to issues of nature and nurture or to currently moot debates, such as whether the nervous system is spontaneously active. Gallistel's great service was to point out the common themes in these classic works and to show how these themes are relevant to current research. There are arguable deficiencies in this book. Not everyone will be pleased by his choice of articles, or by the inconsistent editing style. There were important omissions, particularly in the analysis of organization of behavior at level 5 and above (e.g., McFarland 1974) and in his nearly exclusive focus on locomotion as an example of motor coordination and control. In addition, his conceptual framework will demand considerable work before it can be taken as adequate to account for current data. Gallistel's point that hierarchies are latticelike (that is, individual lower units can be controlled by many different higher units) is well taken, but his illustrations from the military and business suggest that hierarchies are shaped like pyramids, with fewer and fewer units at each successive level. It seems to me that hierarchies in an animal are as often shaped like hourglasses, diamonds, or even upside-down pyramids. Another point: though Gallistel treats reflexes, oscillators, and servomechanisms primarily as low-level units, they strike me as representative functions and relations that are likely to be found at both high and low levels of a hierarchy. Further, I suspect that there are more complex units underlying behavior (Tinbergen 1951; Timberlake 1981), and that levels in a hierarchy may not be fixed but may actually reverse under different circumstances. Last, any assumption of ascending levels of control culminating in the neocortex makes me suspicious of a species bias. Gallistel's slightly cautious interpretation of the transection data provoked some wonder that invertebrates, even fish, reptiles, and birds, manage to get about in an orderly fashion. I doubt that Gallistel would disagree that his scheme is provisional and incomplete. His attempt to include cognitive psychology is suggestive of the potential of a general structural approach, but it did not seem to me entirely successful. This may be more the fault of the dominant models in cognitive psychology than of a flaw in his approach. What he has shown is that we must treat behavior as an entity of complex causation and control. In so doing we may be forced to combine the varying conceptions underlying our research into a common final path, grounded in behavior but adequate to the complexities of information processing and decison making. The nature of learning. It may be fervently hoped that one effect of Gallistel's integration will be to drag our conception of learning into the twentieth century. Until recently, the study of animal learning has suffered gravely from an impoverished model of the functioning organism, grounded in a simplistic and misinterpreted concept of the reflex that was common at the beginning of the twentieth century. (Gallistel's book is worth the price if only because it quotes enough of Sherrington to show that a functioning reflex was not a simple automatic unit of behavior, but a complexly controlled, integrated, and adaptive entity.) In most traditional studies of learning, an animal is assumed to bring to the test chamber only basic sensory and motor equipment, a few reflexes, some regulatory tendencies, and its past learning. Even this simple equipment was denied in part by those who attempted to reduce regulatory tendencies to an associative basis (Hull 1935; see Timberlake 1980), and reflexes to learning occurring in ontogeny (see Kuo 1967).

The education of behaviorism and the nature of learning

William Timberlake
Department of Psychology, Indiana University, Bloomington, Ind. 47401

The education of behaviorism. At last! Someone in the mainstream of psychology has assembled in one book much of the historically important work on behavioral control, and attempted to integrate it into a general scheme that accounts for current research. The result is a breath of fresh air for teachers and researchers trying to develop a more adequate conception of behavior than that provided by the reflex, the operant, and neural pathways and centers. Most of us knew of other conceptions, but reading in secondary sources about taxes, oscillators, servomechanisms, and hierarchical control

638

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

Response/Gallistel: Organization of action

Such oversimplification focused all attention on the hypothetical strengthening process that presumably intervened between simple sensory reception and motor output to produce learned behavior. This focus isolated the study of learning from the animal's ecology (Timberlake 1981), its regulatory processes (Timberlake 1980), and the physiological coordination of its behavior. In short, most of us have been studying learning using positivistic assertions that certain concepts exist (because we have developed traditional methods for measuring them), combined with outmoded and incorrect reflex physiology (or models of learning derived from that physiology). It is time to make the study of learning more continuous with the study of motor coordination and control. It is damaging to each approach to maintain sharp distinctions between them. The physiological units of motor coordination are vitally involved in learned behavior, and in many cases learning is no less involved in the control and integration of motor output. The investigation of phenomena such as bird-song learning, with its clear intertwining of physiological mechanisms and learning, may well be fundamental to the understanding of all learning. A potential contribution of researchers in learning may be to develop more complex models of the structures an animal brings to the learning situation. I recently suggested such a model to account for the behavior of two rodent species toward conspecifics that predicted food (Timberlake 1981). Within this model, low-level sensorimotor units (such as reflexes, taxes, and oscillators) are combined into larger units of perceptual-motor organization called modules. Each module consists of responses that show statistical sequential/ temporal relations and can be elicited, controlled, and terminated by particular stimuli. The modules are organized in a loosely hierarchical fashion into systems serving a common function (such as the feeding system). Potentiation consists of input at the system level, and depotentiation depends on the nature and temporal relation of controlling stimuli. Learning is presumed to occur within these systems in the form of modification of the frequency, order, timing, integration, and elicitation of subunits, modules, and systems. Further, depending on the level and function, learning may occur as the result of response feedback, stimulus-response pairing, motor repetition, or stimulus presentation. The empirical realization of such a general model is a monumental task. It may well be no more than a heuristic guide to other models. One hopes that Gallistel's timely integration will facilitate the development of better models at the interface between research in motor coordination, control, and learning. ACKNOWLEDGMENT I thank Connie Mueller for his conversation.

propose for analyzing the action hierarchy and the nature of representations and how they are used by the action system. I am glad that these themes dominate, because they are the dominant themes of the book itself. I have grouped my responses under three headings: "Fundamental principles for analyzing the structure of action," "Learning and representations," and "Miscellanea." Under "Miscellanea" come the more or less inevitable claims that nothing new has been said and the equally inevitable complaints that I did not fry someone else's favorite fish. Also included in this category are responses that primarily raise interesting and pertinent questions to which I have no good answer or that voice sentiments with which I basically concur.
Fundamental principles for analyzing the structure of action

Bolles. I agree that the essence of the hierarchical notion is that there are units capable of carrying out certain functions on their own and then there are superordinate units that integrate these functions to carry out more complex functions. The "autonomy" of the lower units, the fact that they have a structure that enables them to execute certain operations (e.g., stepping a leg) in the absence of any intervention from above, is of primary importance. This is why I am inclined to doubt the view that in the higher animals every kind of operation is carried out in the cortex (cf. Hollerbach). The kinds of "feedforward" or anticipatory correction mechanisms that Bolles calls attention to in his hand-lifting example are clearly important. I did not discuss these because I am not aware of any principled analysis of how they are achieved. The few discussions I have seen in the AI (artificial intelligence) literature strike me as ad hoc, contrived solely for the purpose of explaining the example at hand, without any concern for generalizability or even testability. Perhaps, as Hollerbach seems to suggest, each scheme is unique to a particular case; if so, we are going to have a hard time creating our science. I am inclined to think that there must be some generalizable principles involved. The principles described by Dev seem to me very promising. Chappie. The questions Chappie raises are the kinds that need to be asked again and again in the study of behavior. Is this particular piece of behavior an elementary or a complex unit of behavior? If elementary, what kind of elementary unit? Are there elementary units other than reflexes, oscillators, and servomechanisms? If so, what are their distinguishing characteristics? If we have to do with a complex unit, what are its constituent units and how do they interrelate? Let us take the so-called scratch "reflex" as an example. Graham Brown (1910) showed that it has two constituents, a reflex raising of the paw to the point to be scratched and an oscillatory scratching motion. Therefore the scratch "reflex" is not an elementary unit of behavior. It is a complex unit, because it can be analyzed into at least two constituents, each of which has the essential features of a unit of behavior initiators (the sensory receptors that drive the reflex
639

Author's Response
Matters of principle: Hierarchies, representations, and action
C. R. Gallistel
Department of Psychology. University of Pennsylvania, Philadelphia. Pa. 19104

The two dominant themes in the BBS multiple review of my book are the adequacy of the principles I

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

Response/Gallistel: Organization of action positioning of the limb and the pacemaker that drives the rhythmic scratching), conductors, and effectors. Since the scratch "reflex" is not an elementary unit, it should not be called a reflex, a name I would have us reserve for use in describing a particular kind of elementary unit. This is the kind of principled analysis that I think we should try to bring to the understanding of behavior. Does one see examples of heterarchical organization at lower levels? Probably, but I think that the generalization that this kind of organization is more common at higher levels is on safe ground. We will only be secure in making these kinds of generalizations when we have a great many very detailed analyses of behaviors at all levels. Unless we strive always to make these analyses in terms of clearly stated principles, we will not build a strong conceptual edifice. So, if there are other kinds of elementary units (and I do not doubt that there are), let us set about stating their distinguishing characteristics, as I have tried to do for the reflex, the servomechanism, and the oscillator. If there are other important principles by which higher levels control lower levels (and I am sure there are), let us try to find concrete examples of them, as I have tried to find concrete examples of control by selective potentiation and depotentiation. Craske. One of the main themes of the book is indeed a critique of those who think of behavior only in terms of reflexes or only in terms of reflexes and voluntary behavior, operant behavior, etc. I agree that Craske's observations suggest oscillators as significant components of everyday movements in humans. Had I been aware of his work, I would have cited it. Dev. As Dev no doubt knows, within any finite interval, a good approximation to an aperiodic trajectory may be produced by a Fourier series that is not only finite but in fact of modest length (<100 terms). However, I am inclined to agree that aperiodic trajectories are not planned in Fourier terms. (I am by no means confident that even periodic movements are planned in this way.) Dev's comments highlight an important concept that I failed to cover in my book - namely, the idea that the controlled variable in many (maybe all) movements is the set of length-tension curves for the muscles operating on a joint. For any particular limb position, there is a family of sets of length-tension curves. Each member of the family is a set of curves, one for each relevant muscle. The intersection of the curves in a set yields the desired position, subject to an error determined by the load. The family for a given position consists of all those sets with the same intersection. The higher the average tension in a set belonging to this family, the greater the stiffness of the limb, hence the less the error produced by any given load. The notion that the CNS selects a family based on the particular position it wants, and a set within the family based on the load it anticipates, goes a long way toward explaining how the CNS overcomes the nonlinear mechanical characteristics of limbs in realizing preplanned trajectories. I believe it was Feldman (1966) who first advanced this explanation. The work of Bizzi, Dev, Morasso, & Polit 640
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

(1978) and others (e.g., Cooke 1979; Schmidt & McGown 1980) has provided experimental support for it. As Dev notes, this is a nonservo mode of control. This length-tension (or mass-spring) theory of limb control does not, however, address itself to the question of how the to-be-executed trajectory is represented in the first place. The question is, what is the most natural, economical "language" in which to formulate trajectories? Perhaps most trajectories are specified solely in terms of their end points, with the control of dynamics being limited to the specification of stiffness. Or perhaps, as Dev suggests, selected aspects of the dynamics (e.g., peak velocity) are represented in advance. In that case, any movement will be planned in terms of a modest number of parameters. Obviously, we need experimental data on the question of the extent to which the dynamics of a movement are subject to planned variation. If experiments should reveal that the complete time course of a movement may be mapped out in advance, then -we will have to confront the question: how does the CNS decompose arbitrary three-dimensional vector functions of time into a finite set of elementary functions, each having a finite set of parameters? One such decomposition is Fourier decomposition. As Newell points out, there are many others. For example, aperiodic movements could be planned in terms of a series of step functions. Yet another question that remains to be addressed is the one raised by Bolles. Most robots are anchored in the upright position, but the animal body is not. How does the CNS take account, in advance, of the effect of an arbitrary limb displacement upon postural stability? Does it use a model of the body's biomechanics? On the face of it, such a model would be stupefyingly complex. Are there any simplifications possible? Is there evidence that the CNS exploits these simplifications? We are in no danger of running out of hard questions anytime soon. Doty. I am pleased that Doty finds the first and largest part of my book congenial. I only wish I could convince him and other neurophysiologically oriented physiological psychologists that abstract models with, for the moment, no neurophysiological content are not mere nominalizations. Neurophysiology-free models are indispensable, both for understanding the behavioral phenomena and for understanding the neural basis of those phenomena. As late as the mid-1980s, A. V. Hill remained scrupulously agnostic as to the possible physio-chemical embodiment of excitation in nerves. He treated excitation as an abstract, physically mysterious process. He used data from the nerve-muscle preparation to create a mathematical characterization of this abstraction. The characterization did not specify the physical nature of excitation, but it put strong constraints on its possible embodiments (Hill 1932; 1936). Similarly, the Hodgkin & Huxley model gives a nonphysical treatment of the neuronal membrane, in that it leaves the biophysics unspecified. What is this mystical talk about "gates"? the biophysicist is apt to sneer. But the Hodgkin & Huxley model is not so much hot air; it is a very powerful theory, at its level of analysis. To return to the example I used in the Precis, the

Response/Gallistel: Organization of action molecular genetics revolution is literally inconceivable in the absence of the conceptual framework provided by classical genetics. The possible molecular embodiment of the gene in classical genetics was at least as mysterious as is the possible neurophysiological embodiment of "a node representing food." But that did not make classical genetics a collection of empty words papering over ignorance. I think Deutsch's (1960) theory is wrong. My Fourier models probably are too. But they are not mere words. They give fairly explicit directions for constructing a machine that exhibits the phenomena that these models seek to explain. The "principle of response generalization" is, by contrast, a mere nominalization. This so-called principle gives no indication of how one would build a machine that exhibited the phenomena that Wickens (1938; 1939) found in his conditioned finger withdrawal experiments, or any of the other phenomena psychologists discuss under the heading "response generalization." The same is true for the principle of stimulus generalization and most of the other principles in the impoverished arsenal of S-R psychology. Fentress. I think the kinds of worries Fentress has about the notion of units in behavior apply with equal force to the units in other scientific areas. The atom, the proton, and the molecule are all units, but they are none of them perfectly stable, nor indivisible, nor free from complex interactions with other units. That an atom is divisible into electrons, protons, and neutrons does not make it any the less a unit in scientific analysis. That oxygen and hydrogen can combine to form water, a substance with properties markedly different from those of hydrogen and oxygen, does not mean that hydrogen molecules and oxygen molecules are not units in chemistry. That the radium atom is not stable does not make it any less a unit. Nor do these complexities mean that one should abandon the notion of units in chemistry and physics and talk only of process. Talk about a process is intelligible only when one has well-defined units that enter into the process. None of this should be taken to mean, however, that I think that an adequate concept of the kinds of units we have to deal with in behavior and the kinds of interactions these units enter into is easily arrived at. The examples Fentress gives nicely illustrate the kinds of problems we have to come to terms with. Grillner. I am pleased that Crillner feels I have adequately captured the general picture of movement control held by those doing experimental work on the mechanisms of movement. This was my primary goal. In retrospect, I see that the fatal word "new" in my title ought to have been left out. This bit of puffery on my part clearly raised some hackles. The unstillable voice of demon pride still whispers in my head that, while these ideas have been widely held, no one has put them all together in one place in recent years, building from muscles to motivation and representation. But enough of this vanity; indeed these ideas are common currency, both within the experimental community, which I knew, and, it appears, within the artificialintelligence community, which I had not realized. The task now, as Crillner, Hollerbach, Lewis, and others urge, is to look ahead to a lengthier and more recondite work that puts flesh on these bones, by focusing on modern experimental work in this active and exciting field. Hogan. By raising Dawkins's (1976) distinction between hierarchies of embedment and hierarchies of connection, Hogan brings an important issue into sharp focus. I agree that at the behavioral level of analysis the system I describe is a hierarchy of embedment. The stepping of an individual leg is embedded in walking. Walking is embedded in approaching food. Approaching food is embedded in nutritive behavior, and so on. What we find at the neural-circuit level of analysis may be viewed as either a hierarchy of connection or a hierarchy of embedment. There is an intraganglionic neural circuit whose function is to control the muscles that cause an individual leg to step. This circuit can be made to operate all by itself in the isolated hemiganglion of a roach. Distinct from and superordinate to the intraganglionic leg-stepping circuits is the interganglionic circuitry whose function is to interrelate the six leg-stepping circuits so that the legs step in a walking pattern. This interganglionic circuitry carries out a higher-level function. It does so by regulating the operation of the intraganglionic leg-stepping circuits. The relation between the interganglionic and intraganglionic circuits is a hierarchy of connection. On the other hand, when we shift our attention to the neural circuitry that interrelates the operation of walking circuitry and orienting circuitry to produce approach, we may regard the walking circuitry as a unit, in which case the circuitry for stepping an individual leg is embedded in the walking circuitry. In short, intermediate units of behavior are not all at one level, as Hogan suggests. There are many levels. Intermediate units at higher levels have intermediate units at lower levels embedded in them, whether one chooses to work at a behavioral or at a neural-circuit level of analysis. I agree that the pyramidal neuron route from the cortex directly to the motor neurons is not part of the diencephalon-dominated hierarchy that I have described. I suspect that this extrahierarchical route to the motor neurons is used by higher vertebrates for the kind of special-purpose direct-access programming that I allude to in my response to Hollerbach. In computers called on to perform special computations at great speed, the usual hierarchically structured highlevel programming languages do not allow the programmer to write the most efficient program. In such cases, machine-language programmers are called in to write a maximally efficient program in the lowest-level language. The pyramidal tract may be the agent of machine-language programming systems in the cortex (another outrageous speculation, I fear). There is no evidence that I know of that the waving of a fish fin can be analyzed into constituents that are units of behavior in their own right (whereas there is such evidence for the scratch "reflex"; see my response to Chappie). Until there is such evidence, this unit should be regarded as an elementary unit of behavior, not a complex unit, which is what all units at Level 5
THE BEHAVIORAL AND BRAIN SCIENCES (1981). 4

641

Response/Gallistel: Organization of action are. The same can be said for the vestibulo-ocular reflex, the optokinetic servomechanism, the flexion reflex, the proboscis extension reflex in the blowfly, the short-latency tail-flip reflexes in the crayfish, and many other examples. In short, I think there really are elementary units of behavior, and there really are complex units of behavior, and there are principled procedures for telling which is which. Hollerbach. This review raises a great many interesting points. I am afraid I have neither time nor space to respond adequately, but here goes. Yes, I subscribe to Church's thesis. My description of oscillators, servomechanisms, and reflexes is indeed coarse-grained and does not do justice to the richness of the conceptual and methodological material that one encounters upon delving more deeply into any of these topics. My choice of older, simpler papers and a relatively simple level of discussion was determined by the intention to provide students with an initial grasp of these entities and to give them some concrete illustrations of how these units combine to produce more complex behavior. I believe in the historical approach to the understanding of major ideas, scientific or otherwise. Plunging students straight into the latest work causes all but the best to lose the forest for the trees. Modern work on servomechanisms makes heavy going for students who have largely forgotten whatever calculus they may have learned. You have to motivate their plunge into frequency domain analysis, integral control, predictive control, etc. by first giving them some feel for how all this might add up to an understanding of behavior. I am not sure what to make of the comments on the concept of the reflex. They seem to imply that I regard oscillators and servomechanisms as reflexes, whereas I was at pains to distinguish them from reflexes. The vestibulo-ocular reflex is a reflex. The optokinetic reaction is not a reflex; it is a servomechanism. The muscle spindle system, if it in fact functioned as Merton (1973) imagined it might, would be a servomechanism; but, as Hollerbach notes, it does not seem to function as a load-compensating servo. The long-loop compensation mechanisms now enjoying favor cannot be treated as simple servomechanisms, for just the reasons Hollerbach cites. That a reflex is subject to potentiation and depotentiation from above does not make it any the less a reflex. Sherrington knew better than anyone about the potentiation and depotentiation of reflexes; he showed the dramatic effects of various kinds of brain transections upon the elicitability of spinal reflexes. Sherrington's dualism, like Newton's theology, seems to have had little impact on the essentials of his scientific thought. He did not think that these transections affected the actions of the soul, although he apparently did think that any transection separated the neural machinery from the influence of the soul. In my mind the jury is still out on how bad frequency domain (Fourier) theories are in vision and audition. Are time and space domain theories noticeably better? It seems to me that Hollerbach's (1981) own work on handwriting encourages one to think of the representation of these movements in Fourier terms; but I would be the first to admit that no
642
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

compelling case has yet been made. I agree that feedforward control is clearly important (cf. Bolles), and I regret that I did not include some material on this in the book. I think that the nervous system of advanced vertebrates has indeed found it necessary for some purposes to give very high levels direct access to very low levels. But I doubt that the already evolved intermediate units of coordination, which function well in most routine situations and which need so little instruction from above, have been abandoned in favor of a system in which the cortex decides what every muscle is to be doing at every moment. I agree that the heterarchy-hierarchy distinction is a rough cut, but a useful one nonetheless. By insisting on the hierarchical character of the overall organization of the brain we remind ourselves that there are higherlevel functions that are realized by the appropriate control of lower-level functions. And we remind ourselves that the behavioral consequences of an intervention, be it instruction, brain stimulation, lesions, or what have you, can only be understood if one has an idea of the level of functional integration at which the intervention produces an alteration. Knowing the level does not make one's understanding in any way complete, but until one knows the level, understanding has hardly begun. By acknowledging the presence of heterarchical organization we remind ourselves not to force everything into the hierarchical framework. One often encounters modes of interaction that do not fit the framework, and one has to be alert for them. However, I agree that as we acquire a better knowledge of the control structures that solve particular kinds of behavioral problems, we will need a much richer set of categories for classifying these structures. Like Hollerbach and von Hoist, I think the attempt to understand behavior in terms of a very few primitives has been a failure. We have slit our throats with Occam's razor. However, I hope we do not now go to the opposite extreme and imagine that the control structure for each piece of behavior is unique, working by principles not found in any other control structure. Each structure may offer a unique combination of principles. And some structures will prove to employ unique principles. Most structures will prove to employ principles also used in other structures, at least if Einstein was right in assuming that God is subtle but not malicious. Jander. I am disappointed that my book has received so harsh an evaluation from a quarter where I hoped it might find favor. Jander (1957) did seminal work on the mechanism of menotaxis (for follow-ups, see Linsenmair 1969; 1970; 1973). Franklin, Bell, & Jander's (in press) work on the leg movements made by the cockroach in turning is an important new departure in the analysis of the locomotory control structure. The works just cited are fine examples of the detailed behavioral analysis of control and coordination mechanisms, the kind of detailed analysis that several reviewers call for and that I hope my book promotes. Such analyses are most useful when accompanied by a careful attention to principled distinctions. Jander's criticisms give me occasion to reemphasize some of the

Response/Gallistel: Organization of action principled distinctions I make in my book. First, I used the term elementary units of behavior advisedly because I explicitly include (not exclude, as Jander recommends) the input side. I follow Sherrington in distinguishing between a unit of action, such as the motor unit, which has no initiator (see Precis), and a unit of behavior, such as the reflex, which does. Since most units of behavior have a sensory/perceptual component, it would be madness to build a theory of action on units that are in principle denied such a component. I defined my units in the way I did precisely in order to avoid the inconsistency that Jander imputes to me. I give explicit principles for distinguishing between complex units and elementary units and for distinguishing my three kinds of elementary units into mutually exclusive classes (see Precis for a recapitulation of these principles). The categories of elementary units, while mutually exclusive, are not exhaustive, because I think there are other classes yet to be defined. I do not think there are only three classes of elementary units and I explicitly say so (Organization of Action, henceforth Action, p. 210). Reflexes are common, and they are not servomechanisms or parts of servomechanisms. The vestibulo-ocular reflex, saccadic eye-orienting reflexes, and the flexion reflex to painful paw stimuli are a few among innumerable possible examples of elementary units of behavior that are reflexes and not servomechanisms. Some units that are sometimes called reflexes in common parlance are in fact servomechanisms (e.g., the optokinetic reflex, which I am careful to call the optokinetic reaction). I hope that my book will serve to correct some of these inaccuracies in common parlance by focusing attention on the need for a principled terminology. The feedback component is irrelevant in saccadic eye-orienting movements because altering the feedback during the course of the movement has no effect on the course of the movement. Ditto for the flexion reflex. This is the principle by which one distinguishes reflexes from servomechanisms in the many cases where the possibility of servomechanistic feedback exists. I clearly state this principle in Action (p. 145). Where is the arbitrariness in any of this? This is a textbook prescription for distinguishing servo control from nonservo control: open any loops and see if that matters. If my principles are adhered to, there is no danger of classifying feeding behavior as a servomechanism, because feeding behavior is obviously a complex unit of behavior. The term "servomechanism" applies only to elementary units. There are, of course, servomechanistic aspects to the control of feeding behavior, as I point out in chapter 10, but the theory of simple servomechanisms is transparently inadequate to capture these. The scratch "reflex" is indeed a complex unit of behavior, one of whose constituents is an oscillator (see my response to Chappie). Again, if my principles are adhered to, there is no danger of misclassifying this unit. I tried to adhere to the distinction between an oscillator, which is a unit of behavior, and a pacemaker, which is the autonomously active neural element or circuit that serves as the initiator constituent of an oscillator. I may have slipped up on this in one or two places; if so, I am sorry. My use of "taxis" generally follows Fraenkel & Gunn's (1961) distinction between the orienting mechanism (the taxis proper) and the locomotory mechanism, whose operation may or may not accompany the operation of the orienting mechanism. The violation of this usage in Action, p. 275, is implicit, not explicit. It occurs in a passage that refers back to Fraenkel's paper (reprinted in chapter 6), a fifty-year-old paper in which this distinction was not made. Jander's going to this length to find fault seems to me a little bizarre. The same applies to his criticism of my definition of an elementary unit of behavior in the Glossary. Jander omits from his quote the next sentence of the definition, which begins, "It [the circuit] must have a component or components in which nerve signals arise. . . . " That makes clear provision for the sensory receptors in reflexes and servomechanisms, a point on which I cannot imagine anyone reading my book being in any doubt. The fault Jander finds here would seem to apply only if one follows his (not my) proposal to disregard reception, perception, and cognition in building a hierarchic system of concepts to describe action. At several places in my book I call attention to the fact that the lattice-hierarchical arrangement requires intralevel mechanisms that decide which of the many competing units actually has control over lower-level units at any one time. In my discussion of recurrent reciprocal inhibition (Action, pp. 60-63), I discuss one such mechanism at length. How I can be accused of ignoring this problem baffles me. To say that this problem falsifies the lattice-hierarchy theory does not make sense. I find the remarks on my Fourier models incomprehensible. The magnet effect, far from being inconsistent with such models, may be invoked to explain how the oscillators maintain the proper phase relationship when the pattern is executed. That some processes in the nervous system are nonlinear does not mean that they all are. I agree that data on navigation after detours require a theory of spatial representation different from Deutsch's, and I say so in the book. I propose calling the framework espoused in my book "the neuroethological theory of action (in recognition of its roots in behavioral neurobiology and ethology)" [Action, p. 361]. I pay a great deal of attention to the work of von Hoist, Lorenz, and Tinbergen. I cannot imagine why Jander seems to feel that I follow the psychologists' long tradition of ignoring or rejecting ethological theories. On the contrary, I like to think of my book as a modern synthesis of the ethological position on the structure of behavior. Mittelstaedt has written to me that he plans to respond to my critics in a later round in this journal, rather than writing anything for this round. It will be interesting to read his response to Jander, since Mittelstaedt writes from the citadel at Seewiesen. Jeannerod. I find myself in agreement with Jeannerod and with the point made by his interesting experiment. His arguments illustrate the futility of attempting to build a theory of action with units that are denied an
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

643

Response/Gallistel: Organization of action input component. Rather than banning input from consideration, as Jander recommends, I ought to have devoted more attention to it, as Jeannerod recommends. Lewis. I agree that dynamic considerations are central to understanding how sophisticated servomechanisms really work. I mention in my Precis (but not in the book) the fact that stability considerations require a poor high-frequency response in the optokinetic reaction. I heartily second Lewis's recommendation that the serious student delve into more recent literature, where much more sophisticated treatments of servomechanisms are to be found (e.g., Baarsma & Collewijn 1974; Carpenter 1977; Crago, Houk, & Hasan 1976). The student who does so should be forewarned, however, that a working knowledge of linear systems theory is essential (see Caelli 1981; Carpenter's Appendix; Norman 1981; and Riggs 1970, for biologically or psychologically oriented introductions to linear systems theory). Nashner. I am pleased that Nashner finds my exposition of the principles of motor organization congenial. I think his own work provides excellent examples of learned selective potentiation and depotentiation at work in human motor control (Nashner, Woollacott, & Tuma 1979). I regret that I became aware of his work too late to include some of these examples in the book. falling within the classical-conditioning framework, a conceptual edifice about which I am skeptical. In my heart of hearts I do not believe that the process of classical conditioning, as traditionally conceived, ever occurs; nor, for that matter, does the process of instrumental conditioning. I think the classic conceptions of the processes underlying these experimental phenomena are simply wrong. If we relax our conception of classical conditioning to the point where it refers to the learning of any stimulus relation, no matter what the relation, no matter what the mode of representing that relation, and no matter what the procedures for translating the representation into action, then of course the conception applies. But, in this relaxed form, the "theory" of classical conditioning begs all of the interesting questions. It disguises rather than highlights the diversity of plastic phenomena. It avoids raising all of the difficult questions about the nature of representations (see Miiner and Reed). It takes no account of the different levels of organization and the differing consequences of alterations at each level (see MacKay). Miiner. Since in Miiner's case my critic directs his first fire at von Hoist and Mittelstaedt's paper, I'll leave the return salvo largely to Mittelstaedt (who, as I mentioned, has written me that he plans to respond to my critics in a later round). I note only the following: von Hoist and Mittelstaedt do not pretend to offer a complete theory of movement perception. Perceptions of movement in train stations seem to me irrelevant to the issue of how false reafferent movement signals are dealt with. Despite our extensive experience that buildings and trees don't move, the reader can easily verify the fact that when you push on your eyeball with your finger, the whole visual scene, including trees and houses, does seem to move. And the whole visual scene is seen to move by subjects who attempt to move their eyes voluntarily when they have a drug- or diseaseinduced debility in the peripheral neuromuscular apparatus for eye movement. Together, these two observations seem to me strongly to imply the operation of a mechanism that algebraically cancels reafferent visual-field-motion signals with a copy of the signal that normally commands eye movement. As regards our realizing that something is no longer in the line of gaze after an eye movement, this objection presumes that perceptions of movement are not mediated by a separate channel from perceptions of spatial locus; whereas there are both behavioral and neurophysiological reasons for believing that these percepts are handled by separate channels. I think it is highly unlikely that our movements through space are directed by a code that explicitly specifies the outcome of each motor performance. Naturalistic observations by Calhoun (1963), the experiments by Maier (1929) and Thorpe (1950) reviewed in my book, and a great deal of material reviewed by O'Keefe & Nadel (1978) [See BBS 2(4) 1979] show that animals are capable of taking a direct route from any point in a familiar environment to any other. The digger wasp transported to an arbitrary point finds its way to its unseeable burrow. The triggered motor program (S-R) view advocated by Miiner requires an infinite number of learned programs and even then

Learning and representations

MacKay. I agree that I slighted Lashley, but not because I don't admire him. His work just did not happen to fit conveniently into the exposition. I do, however, explicitly acknowledge his early advancement of the selective-potentiation view of motivation. Also, his work is very well known to psychologists. I was trying mostly to bring together less well known work pertinent to some of the problems Lashley highlighted, though not, I admit, to problems of morpheme and phoneme sequencing in speech or to the problem of finger sequencing in the playing of an instrument (classic problems in the serial order of behavior, about which I simply have nothing interesting to say). I also concur with MacKay's argument that an abstract muscle-independent representation of learned movements such as handwriting is not the whole story. There must indeed be effects of experience at other levels, effects that have a considerable bearing on the fluidity and precision of the process that translates the muscleindependent representation into muscular expression. In my desire to show that one could give a concrete neurobiologically plausible model of the abstract representation, I did not emphasize enough the importance of plastic alterations at lower levels of the hierarchy. MacKay and I are one in believing that plasticity and learning cannot be intelligibly discussed without taking into account the different levels of the hierarchy and the differing functional consequences of plastic alterations at different levels. Mackintosh. I do not see my notions of learning as
644
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

Response/Gallistel: Organization of action cannot explain how the animal sets an appropriate homeward course from a point it has never been to. Setting a course from a point where one has never been toward a home (or other goal) that gives off no stimulus beacon detectable at a distance requires reference to landmarks that define a spatial locus for both the goal and the starting position. This reference presupposes a map that localizes the goal in a space defined by landmarks, at least some of which are detectable from the starting point of the course. The only possible alternative I am familiar with - inertial navigation relies on double integration of the acceleration vector. Since the capacity to find one's way home without relying on beacons or trails seems very widespread in the animal kingdom (cf. Jander), I believe that the ability to make and use spatial representations is equally widespread. Either that, or animals have inertial navigation mechanisms of heretofore undreamedof precision. It is true that Deutsch's theory offers no clear indication of how nonspatial representations might be translated into appropriate actions. I think his theory is inadequate to deal with even spatial representations (though it is useful for illustrative purposes, as Milner notes). However, I think a more adequate theory will be even more peculiarly spatial. I think the quest for a general theory of learning of the kind Milner wants is a will-o'-the-wisp that has sidetracked psychology for most of this century. There are, I believe, many distinct organs for learning in the brain [cf. Chomsky: "Rules and Representations" BBS 3(1) 1980], each prepared to learn about quite different aspects of reality. There are equally as many readout systems for translating what has been learned into action. The organ (brain circuitry) that learns where things are is, I believe, both functionally and morphologically distinct from the organ that learns what causes what. And the system that translates spatial knowledge into action is likewise distinct from the system that translates causal knowledge into action, although both systems undoubtedly employ the same hierarchically organized units of action. This pretheoretical bias about the nature of learning and representations raises several perplexing questions: 1. What are the fundamental modes of representation? For many years most psychologists have assumed that there was one and only one fundamental mode of representation, a representation in terms of associations. I am inclined to believe that this mode of representation never occurs. I am inclined to argue that an association, as classically conceived, never forms in the brain of any animal, man included. What do form, I would assume, are among other things, representations of spatial locus. These representations of spatial locus could themselves be represented in a formal psychological theory by sets of Cartesian coordinates. In short, I believe the brain literally makes maps. But formal systems suitable for representing spatial primitives (points, lines, etc.) appear hopelessly ill suited to represent other aspects of reality, such as causation. Therefore, I must assume that there are other equally fundamental modes of representation in the brain. What are they? Good question. 2. How do we decide whether a proposed mode of representation really does exist in the brain? I have nothing to add to what Chomsky (1980) had to say on this difficult question in the pages of this journal. We cannot decide a priori. We decide ultimately on the basis of how coherent and powerful the various proposals prove to be when confronted with an array of experimental and observational facts. 3. Within a mode of representation, what is represented explicitly and what is derived from the directly represented information when the occasion demands? This is the question most directly at issue in Milner's and my opposing views about the nature of the representation underlying navigational competence. Milner assumes that the animal explicitly represents the fact that proceeding X meters from A at an angle a to the line AB leads to the appearance of C; Reed seems to think that this is what I assume. On the contrary, I assume that the animal represents explicitly only the coordinates of A, B, and C within some common spatial frame of reference. From these coordinates the animal derives the required angle and distance of progression when and if the occasion demands. How does one decide who is right? By the kinds of arguments given above, I think. My proposal explains how animals can derive a course from an arbitrary, previously unvisited point, provided only that they are given sensory information that enables them to determine the coordinates of that point within the common frame of reference. Milner's proposal, so far as I can see, cannot explain the ability to set a correct course from an arbitrary point, a point never visited before (assuming always that the point toward which the course is set is not the source of any sensory information at the time the course is set). If experiments confirm that animals can in fact do this, then the Cartesian-coordinates proposal has more merit than the act-outcome proposal. Reed. I agree that most, perhaps all, of the ideas in my book are not new. Many can no doubt be found in some form in the writings of the ancients, just as can the germs of the atomic theory of matter. I think what counts in science is not the propounding of an idea but the marshalling of a compelling body of facts and arguments in support of an idea. By the way, did Descartes really say that the cognitive machinery selectively potentiates the machinery of the body? He said the soul directs the machinery of the body, but did he say it selectively potentiates it? I'm no Descartes scholar, but it doesn't sound like him. In any case, that is not what what I say. I say that selective potentiation by a motivational signal of points in a representation determines which external stimulus will be oriented to. The problem of how representations are translated into action is indeed a vexing one, and I make no claims to have solved it. I object, however, to the characterization of my speculations about the neural mechanisms underlying the brain's representation of spatial configurations as vague and desultory. As explained in the Precis - [not seen by reviewers; ed.] - as well as in the book, the idea is that the distribution of matter in three-dimensional space could be represented in the frequency domain; that is, the brain could use signal vectors specifying the frequencies, amplitudes, phases, and orientations of Fourier components. If the repreTHE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

645

Response/Gallistel: Organization of action sentation does take this form, then it could be rotated by changing the values of the phase and orientation signals in each vector. Such a rotation would necessarily be continuous. These suggestions are speculative in the extreme, "outrageous" even (cf. Hogan), but not vague. Nor are they desultory in a discussion of how the brain might form spatial representations and utilize them in the control of action. I agree that proposals about spatial representations should be designed to generate tests and explain facts. Kenneth Cheng and I are working on a combined formal and experimental analysis of the hypothesis that rats utilize a Euclidean representation of their environment in finding their way about in the dark (see also Landau, Gleitman, & Spelke, in press). There are, however, some experimental facts already at hand, facts that badly want explanation. One such fact is that digger wasps, who make a visual survey of their newly dug burrow, can use the results of this survey to find their way home from any location within about 100 meters away, even from never-visited locations to which they have been transported inside a closed box (Thorpe 1950). Since there is compelling evidence that the burrow itself gives off no stimulus beacon detectable even a meter away (Tinbergen & Kruyt 1938), I know of only two possible explanations: the wasps (1) use a map made during their visual survey, or (2) carry out nearly perfect double integration of the acceleration vector over the better part of an hour while moved this way and that over hundreds of meters, both by .their own efforts and by the whim of the experimenter. I favor the first of these two explanations. If Reed (or anyone else) knows of an alternative explanation, or of any reason to believe that such a feat of double integration is remotely conceivable, then I am genuinely anxious to learn of it. (Solar navigation belongs in the map class, since to find its way home from an arbitrary point using the sun as a reference the animal must know the latitude and longitude of its burrow - and have a wondrously good clock and an astonishingly precise system for shooting the sun.) Which brings me to counterfactual conditionals and diving gannets. The reader should know that my book says absolutely nothing about either of these topics. I would not dream of suggesting that the gannet generates the appropriate wing retractions through the agency of a counterfactual conditional. Like Lee (1980), I assume the retraction is triggered by some aspect of the expanding optical array. How this mechanism has evolved I do not pretend to guess. The fanciful evolutionary reasoning in which Reed indulges on this subject strikes me as the sort of thing that has given evolutionary reasoning a bad name, a bad name much resented by those who make more disciplined evolutionary arguments. Do representations go beyond the information given? Yes; if they did not, they would be of little use. Chomsky (1980) has made this argument better than I can (his "poverty of the stimulus" argument and accompanying references). Let me illustrate, in the spatial domain, why representations must go beyond the information given. A blind girl led from A to B and back to A, then from A to C, can apparently compute the approximate angle and distance she must go to
646
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

reach B from C (Landau et al., in press). The system underlying her movements through space seems to employ trigonometric calculations, with the sensory data obtained on earlier routes used to compute the new route. Sensory data about the distances AC and AB and the angle BAC would make the calculation of the distance CB and the angle ACB possible only if in representing these data the brain assumes that the room is a Euclidean space. If the space were assumed to be Riemannian, there would be no way to calculate the required distances and angles in the absence of information about the local curvature of space. Of course, the space is in fact Riemannian, although the curvature is so weak that the space departs by immeasurably small amounts from the Euclidean within the confines of the room. The point is that if the child's system for spatial representation were to adopt an agnostic stance vis-a-vis the Euclidean or non-Euclidean character of the space, it would not be able to support the requisite calculations, calculations whose biological utility for a blind child (or a rat in the dark) is tolerably obvious. Roitblat. It will come as startling news to Chomsky (1980) that a strong commitment to reductionism and physical realizability permits nothing but a modified behaviorism. Chomsky is strongly committed to reductionism and physical realizability. If his theories are but modified behaviorism, then behaviorism is more protean than was heretofore suspected. I agree that I do not attempt to specify conditions under which learning will occur at one or another level of the hierarchy. My views on the disparate character of different kinds of learning preclude any general attack on this problem. They argue instead for domain-specific analyses. Since there is a lot of discussion of representations in the last chapter of my book, and since the laying down of representations underlies many kinds of learning, it is erroneous to say that the only mechanisms that could account for learning are the selective potentiation of reflexes and servomechanisms and the selective coupling of oscillators. The laying down of a spatial representation has nothing to do with servomechanisms, oscillators, or reflexes, and I do not see how these maps can be viewed as "stored in . . . an action hierarchy." Like Roitblat, I think representational systems are largely orthogonal to the sensorimotor system. Of course, the sensorimotor system must have access to the represented information; otherwise there is no point to the brain's representing information in the first place. Does any theory that grants the action system access to the representation include the knowledge "as part of the action system"? If so, then the library at my university must be seen as part of my action system, as must the road maps I use in driving to unfamiliar places. In any decompositional representation of space that has orientation information directly represented by distinct signals, rotation of the representation will take place by changing the strengths of those signals. It is hardly ad hoc to assume that those changes will be continuous; most signals change strength continuously. Timberlake. After hearing from the AI fraternity that everything in my book is old hat, it is pleasant to find

Response/Gallistel: Organization of action that someone thinks that the synthesis at least is new, which is the farthest I would go in claiming originalityAs regards the supposedly pyramidal shape of my hierarchy, see the section on oligarchs in the Precis. In chapter 10, I explicitly call attention to the servomechanistic, oscillatory, and reflexlike character of some of the processes at the top of the hierarchy. But, as explained in my response to Jander, one must be careful to avoid saying that these complex units at the top are servomechanisms, reflexes, or oscillators. To say that they are is a little like saying that an organic acid is a hydrogen ion just because it is a proton donor. Any acid has some of the characteristics of the hydrogen ion, but if we start calling all acids hydrogen ions, nothing but confusion will result. I do not think I mention the neocortex anywhere in my book. For me, the hypothalamus and the forebrain structures with which it is immediately connected constitute the top of the action hierarchy. Although I do not say so in the book, I imagine that cortical structures are used primarily for analysis, representation, and planning, and perhaps also for specialpurpose interventions in very-low-level operations (cf. response to Hogan). Therefore I am hardly baffled by the ability of fish, reptiles, and birds to get through life. They all have well-developed diencephalons and, evidently, enough forebrain to handle their more limited analysis and planning problems. The vertebrate divisions of the brain, of course, do not apply to invertebrates, but in invertebrates it is clear that higher ganglia subserve higher functions, just as in vertebrates. A cockroach with its head cut off can run, but not to any clear purpose. Needless to say, in the light of the many remarks I have already made on the topic, I concur with Timberlake's criticism of the traditional approaches to learning, which ignore the hierarchical structure of behavior, and which assume that the concept of an association is adequate to deal with the phenomena of learning. The chorus of voices singing that there are no general laws of learning grows louder all the time. To my ears, it is a chorus of angels. problems in animal behavior that interest me, and I am anxious to see what it has to say. Newell. I agree that many of the notions upon which my argument is built have had some currency for a long time, particularly the notion of control by selective potentiation and depotentiation (which Reed traces back to Descartes). I see my contribution here as (1) providing concrete examples of the operation of these principles; (2) emphasizing the diversity of principles needed - instead of, like Hebb, trying to make a single, basically associationist concept do all the work; and (3) calling attention to the importance of some systems (e.g., systems of coupled oscillators) that have escaped the attention of many. While the term "handwave" makes me bridle a little, I heartily agree that more detailed discussion is needed on every topic. I hope the book gives students their bearings before they plunge into more detailed analyses. The questions raised in Newell's penultimate paragraph are all pertinent. I hope the book may help focus attention on them. The Fourier models are put forward to illustrate the potential benefits of nonobvious decompositional representations. Other types of decomposition may turn out to be preferable. Olton. The question raised by Olton - why animals sometimes make intelligent use of their representations and sometimes do not - is an interesting one. I am afraid I have no answer. The fact that by appropriate manipulations one can make all but the highest animals behave as if they have no representations at all has helped keep radical behaviorism alive into the fourth quarter of the twentieth century. The problem of maintaining tight experimental control while at the same time arranging matters so that rats make full use of their spatial representations is one that Kenneth Cheng and I are contending with right now. We wish that we had some clear, readily articulable, and convincing answers to Olton's thoughtful question. All we have at the moment are hunches. Provine. I agree that the theory of the response has been sadly neglected in psychology. Provine is perhaps as surprised as I am to learn from Arbib and Newell that the artificial intelligence community has had a theory of animal action much like the one in my book for all these years. If my book succeeds, however unwittingly, in popularizing within psychology the AI theory of animal action, then experimentalists like Provine, who do fine work on motor systems, will not feel all alone in the wilderness any more. Reynolds. What I wrote about were the mechanisms and principles underlying the patterning of muscular contractions into complex purposive behavior. I agree that I did not cover social behavior, ecology, communication, the evolution of behavior, or species-specificity in the laws of thought. I also neglected quantum mechanics, the fragmentation of genes in eukaryotic DNA, the rise of the aristocracy in the Middle Ages, the lamentable irrationality of the V2, and the transcendence of ir - capital subjects, every one.
THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

Miscellanea

Arbib. Arbib and I clearly have different tastes. Since he does not say what important concepts I missed in the work he wishes I had dealt with, there is little more to say. Maybe the AI people have known all this all along. If so, they have not, so far as I know, applied it to the kinds of problems in animal behavior that interest me. I did not pay more attention to Greene's work because the few papers I read struck me the way much work in artificial intelligence does - very abstract and schematic with little attention to concrete examples drawn from the experimental literature. I thank Arbib for calling Stelmach & Requin (1980) to my attention. I have not had time to look at it yet; but I will. I was also, I blush to confess, unaware of Arbib & Lieblich (1977). It sounds as though this article does tackle the kinds of

647

References/Gallistel: Organization of action

References
Arbib, M. A. & Lieblich, I. (1977) Motivational learning of spatial behavior. In: Systems neurosdence, ed. J. Metzler, pp. 221-39. New York: Academic Press. [MAA, CRG] Baarsma, E. A. & Collewijn, H. (1974) Vestibulo-ocular and optokinetic reactions to rotation and their interaction in the rabbit. Journal of Physiology 238:603-25. [CRG] Bartlett, J. R. & Doty, R. W. (1980) An exploration of the ability of macaques to detect microstimulation of striate cortex. Acta Neurobiologiae Experimentalis 40:713-28. [RWD] Beagley, W. K. & Holley, T. L. (1977) Hypothalamic stimulation facilitates contralateral visual control of a learned response. Science 196:321-22. (CRG] Becker, J. T. & Olton D. S. (in press) Cognitive mapping and hippocampal function. Neuropsychologia. [DSO] Becker, J. T.; Olton, D. S.; Anderson, C. A.; & Brietinger, E. T. P. (in press) Cognitive mapping in rats: The effects of hippocampal and frontal system lesions on retention and reversal. Behavioural Brain Research. [DSO] Bellman, K. (1979) The conflict behavior of the lizard, Sceloporus occidentals, and its implication for the organization of motor behavior. Doctoral dissertation. University of California at San Diego, La Jolla, California. IJCF] Bernstein, N. A. (1967) The coordination and regulation of movement. New York: Pergamon Press. [CRG] Bindra, D. (1978) How adaptive behavior is produced: A perceptual-motivational alternative to response-reinforcement. Behavioral and Brain Sciences 1:41-91. [NJM] Bizzi, E.; Dev, P.; Morasso, P.; & Polit, A. (1978) Effect of load disturbances during centrally initiated movements. Journal of Ncurophysiology 41:542-56. [CRG] Bizzi, E.; Polit, A.; & Morasso, P. (1976) Mechanisms underlying achievement of final head position. Journal of Neurophysiology 39:435-44. [PD] Bohm, D. (1969) Some remarks on the notion of order. In: Towards a theoretical biology, volume 2, ed. C. H. Waddington, pp. 18-60. Chicago: Aldine Publishing Co. [JCF] Bohn, G. (1905) Attractions et oscillations des animaux sous ('influence de la lumiere. Memoires de Vlnstitut General Psychologique 1:1-11. [CRG] Breland, K. & Breland, M. (1961) The misbehavior of organisms. American Psychologist 16:681-84. [CRG] Brown, P. L. & Jenkins, H. M. (1968) Auto-shaping of the pigeon's key-peck. Journal of the Experimental Analysis of Behavior 75:1-8. [NJM] Brown, T. G. (1910) Studies in the reflexes of the guinea pig. II. Scratching movements which occur during ether anesthesia. Quarterly Journal of Experimental Physiology 3:21-52. [CRG] (1914) On the nature of the fundamental activity of the nervous centres; together with an analysis of the conditioning of rhythmic activity in progression, and a theory of evolution of function in the nervous system. Journal of Physiology 48:18-46. [CRG] Carpenter, R. H. S. (1977) Movements of the eyes. London: Pion. [CRG] Carr, H. & Watson, J. B. (1908) Orientation of the white rat. Journal of Comparative Neurology and Psychology 18:27-44. [CRG] Chomsky, N. (1959) Review of Skinner's Verbal behavior. Language 35:26-58. [ESR] (1980) Rules and representations. Behavioral and Brain Sciences 3:1-61. [CRG] Cooke, J. D. (1979) Dependence of human arm movements on limb mechanical properties. Brain Research 165:366-69. [CRG] Cooper, L. A. & Shepard, R. N. (1979) Transformations on representations of objects in space. In: Handbook of perception. Vol. 8: Space and object perception, ed. E. C. Carterette & M. Friedman. New York: Academic Press. [CRG, HRL] Bruce, I. C. & Tatton, W. G. (1980) Sequential output-input maturation of kitten motor cortex. Experimental Brain Research 39:411-19. [RRP] Burt, C. (1962) The concept of consciousness. British Journal of Psychology 53:229-42. [RRP] Caelli, T. (1981) Visual perception. New York: Pergamon Press. [CRG] Calhoun, J. B. (1963) The ecology and sociology of the Norway rat. Bethesda, Md.: U.S. Dept. of Health, Education, and Welfare (U.S. Public Health Service Publication No. 1008). [CRG] Capaldi, E. J.; Verry, D. R.; 4 Davidson, T. L. (1980) Memory, anticipation pattern learning, and transfer in rats. Animal Learning and Behavior 8:575-85. [HLR] Crago, P. E.; Houk, J. C ; 4 Hasan, Z. (1976) Regulatory actions of the human stretch reflex. Journal of Neurophysiology 39:925-35. [CRG] Craske, B. (1981) Programmed aftereffects following simple patterned movements of the eyes and limbs. In: Attention and performance 9, ed. J. B.

Long & A. D. Baddeley. Hillsdale, N.J.: Lawrence Erlbaum Associates (in press). [BC] Darwin, C. (1958) Autobiography (N. Barlow, ed.). New York: W. W. Norton. [ESR] Dawkins, R. (1976) Hierarchical organization: A candidate principle for ethology. In: Growing points in ethology, ed. P. P. C. Bateson & R. A. Hinde. Cambridge University Press. [JAH] Deutsch, J. A. (1960) The structural basis of behavior. Chicago: University of Chicago Press. [MAA, CRG, NJM, PMM] Delcomyn, F. (1980) Neural basis of rhythmic behavior in animals. Science 210:492-98. [BC] Descartes, R. (1649) Passions of the soul. In: Philosophical works of Descartes, vol. 1, ed. E. Haldane & G. Ross. 1911. Cambridge University Press. [ESR] Dilger, W. C. (1962) The behavior of lovebirds. Scientific American 206:8899. [CRG] Doty, R. W. (1951) Influence of stimulus pattern on reflex deglutition. American Journal of Physiology 166:142-58. [RWD] (1975) Consciousness from neurons. Acta Neurobiologiae Experimental 35:791-804. [RWD] (1976) The concept of neural centers. In: Simpler Networks and Behavior, ed. J. C. Fentress, pp. 251-65. Sunderland, Mass.: Sinauer Associates. [RWD] Eccles, J. C. (1980) The human psyche. New York: Springer International. [CRG] Fel'dman, A. C. (1966) Functional tuning of the nervous system with control of movement or maintenance of a steady posture. II. Controllable parameters of the muscles. Biophysics 11:565-78. [CRG] Fentress, J. C. (1972) Development and patterning of movement sequences in inbred mice. In: The biology of behavior, ed. J. Kiger, pp. 83-132. Corvallis: Oregon State University Press. [JCF] (1976) Dynamic boundaries of patterned behaviour: Interaction and self-organization. In: Growing points in ethology, ed. P. P. G. Bateson & R. A. Hinde, pp. 135-69. Cambridge University Press. [JCF] (1980) How can behavior be studied from a neuroethological perspective? In: Information processing in the nervous system, ed. H. M. Pinsker & W. D. Willis, Jr., pp. 263-83. New York: Raven Press. [JCF] (in press) Ethological models of hierarchy and patterning of species specific behavior. In: Handbook of neurobiology: Motivation, ed. E. Satinoff & P. Teitelbaum. New York: Plenum Press. [JCF] Flynn, J. P. (1972) Patterning mechanisms, patterned reflexes, and attack behavior in cats. Nebraska Symposium on Motivation 20:125-53. [CRC] Fodor, J. (1975) The language of thought. New York: Crowell. [ESR] Forman, G. (1981) Epilogue. In: Action and thought, ed. G. Forman. New York: Academic Press (in press). [MAA] Fraenkel, G. S. (1927) Geotaxis und Phototaxis von Littorina. Zeitschrift fur vergleichende Physiologic 5:585-97. [CRG, ERL] Fraenkel, G. S. & Gunn, D. L. (1961) The orientation of animals. New York: Dover. [CRG] Franklin, R.; Bell, W. J.; & Jander, R. (in press) Rotational locomotion by the cockroach Blattela germanica. Journal of Insect Physiology. [CRG] Frisch, K. v. (1967) The dance language and orientation of bees. Cambridge, Mass.: Belknap Press of Harvard University Press. [RJ] Gallistel, C. R. (1980) The organization of action: A new synthesis. Hillsdale, N.J.: Lawrence Erlbaum Associates. [CRG] Ghez, C. & Vicario, D. (1978) The control of rapid limb movement in the cat. II. Scaling of isometric force adjustments. Experimental Brain Research 33:191-202. [PD] Golani, I. (1976) Homeostatic motor processes in mammalian interactions: A choreography of display. In: Perspectives in ethology, volume 2, ed. P. P. G. Bateson & P. H. Klopfer, pp. 69-134. New York: Plenum Press. UCF] Gomer, P. (1958) Die optische und kinasthetische Orientierung der Trichterspinne Agelena labyrinthic Cl. Zeitschrift fur vergleichende Physiologic 41:111-53. [RJ] Greene, P. H. (1964) New problems in adaptive control. In: Computer and information sciences, ed. J. T. Tou & R. H. Wilcox. Washington: Spartan. [MAA] (1967) Models for perception and action. Proceedings of the First Annual Princeton Conference on Information Sciences and Systems, Dept. of Electrical Engineering, Princeton University, 245-53. [MAA] Grill, H. J. & Norgren, R. (1978) The taste reactivity test. I. Mimetic responses to gustatory stimuli in the normal rat. Brain Research 143:281-97. [CRG] Crindley, G. C. (1932) The formation of a simple habit in guinea pigs. British Journal of Psychology 23:12'!'-47. [NJM] Hall, W. G.; Cramer, C. P.; & Blass, E. M. (1977) Ontogeny of suckling in rats: Transitions toward adult ingestion. Journal of Comparative and Physio-

648

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

References/Gallistel: Organization of action

logical Psychology 91:1141-55. [CRG] Hearst, E. 4 Jenkins. H. M. (1975) Sign tracking: The stimulus-reinforcer relation and directed action. Austin, Tex.: Psychonomic Society. [CRG] Hebb, D. (1949) The organization of behavior. New York: John Wiley and Sons. [AN] Held, R. 4 Gottlieb. N. (1958) Technique for studying adaptation to disarranged hand-eye coordination. Perceptual and Motor Skills, 8:83-86. [MJ] Held, R. & Hein, A. (1963) Movement-produced stimulation in the development of visually guided behavior. Journal of Comparative and Physiological Psychology 56:872-76. [MAA] Hill, A. V. (1932) Chemical wave transmission in nerve. New York: Macmillan. [CRG] Hill, A. V. (1936) Excitation and accommodation in nerve. Proceedings of the Royal Society of London, Series B 119:305-55. [CRC] Hinde, R. A. (1959) Unitary drives. Animal Behaviour 7:130-41. [JCF] Hineline, P. N. 4 Rachlin, H. (1969) Escape and avoidance of shock by pigeons pecking a key. Journal of the Experimental Analysis of Behavior 12:533-38. [CRG] Hollerbach, J. M. (1981) An oscillation theory of handwriting. Biological Cybernetics 39:139-56. [CRG] Hoist, E. von. (1937) Vom Wesen der Ordnung im Zentral-Nervensystem. Die Naturwissenschaften 25:625-31. 641-47. [BC, JAH] (1939) Die relative Koordination als Phanomen und als Methode zentralnervoser Funktionsanalyse. Ergebnisse der Physiologie 42:228-306. (For English translation see v. Hoist 1973.) [CRG] (1973) The behavioral physiology of animals and man: Selected papers. Coral Cables, Fla.: University of Miami Press. [CRG] Hoist, E. von & Mittelstaedt, H. (1950) Das Reafferenzprinzip. Wechselwirkung zwischen Zentralnervensystem und Peripherie. Die Naturwissenschaften 37:464-76. (Translated into English by R. D. Martin and reprinted in von Hoist, 1973) [CRG, SG, ERL, PMM] Hoist, E. von & St. Paul, U. von (1963) On the functional organisation of drives. Animal Behaviour 11:1-20. (Reprinted in von Hoist, 1973) [CRC] Hull, C. L. (1935) Special review: Thorndike's Fundamentals of learning. Psychological Bulletin 32:807-23. [WT] Hulse. S. H. 4 Dorsky, N. P. (1979) Serial pattern learning by rats: Transfer of a formally defined stimulus-relationship and the significance of nonreinforcement. Animal Learning and Behavior 7:211-20. [HLR] lersel, J. J. van 4 Bol, A. C. A. (1958) Preening in two tern species: A study on displacement activities. Behaviour 13:1-88. [RJ] Jander, R. (1957) Die optische Richtungsorientierung der Roten Waldameise (Formica rufa L.). Zeitschrift fur vergleichende Physiologie 40:162-238. [CRG, RJ] Jeannerod, M. (1981) Intersegmental coordination during reaching at natural visual objects. In: Attention and performance 9, ed. J. Long & A. Baddeley, pp. 153-68. Hillsdale, N.J.: Lawrence Erlbaum Associates. [MJ] Kamil, A. C. & Sargent. T. D. (1981) Foraging behavior. New York: Garland STPM Press. [DSO] Kavanau. J. L. (1969) Behavior of captive white-footed mice. In: Naturalistic viewpoints in psychology, ed. E. R. Willems & H. L. Raush. New York: Holt, Rinehart and Winston. [CRC] Klapp, S. T.; Grein, D. M.; Mendicino, C. M.; and Koenig, R. S. (1979) Anatomic and environmental dimensions of stimulus-response compatibility: Implication for theories of memory coding. Acta Psychologica 43:367-79. [DCM] Kosslyn, S. M.; Pinker, S.; Smith, G. E.; 4 Schwartz, S. P. (1979) On the demystification of mental imagery. Behavioral and Brain Sciences 2:535-81. [CRC] Kuhn, T. S. (1962) The structure of scientific revolutions. Chicago: University of Chicago Press. [HLR] Kuipers, B. (1978) Modeling spatial knowledge. Cognitive Science 2:129-53. [MAA] Kuo, Z.-Y. (1967) The dynamics of behavioral development. New York: Random House. [WT] Kupfermann, I. 4 Weiss, K. R. (1978) The command neuron concept. Behavioral and Brain Sciences 1:3-39. [SG] Land, M. (1980) Optics and vision in invertebrates. Handbook of sensory physiology, VII/6B. Berlin: Springer. [ESR] Landau, B.; Cleitman, H ; & Spelke, E. (in press) Spatial knowledge and geometric representation in a child blind from birth. Science. [CRG] Lashley, K. S. (1938) Eiperimental analysis of instinctive behavior. Psychological Review 45:445-71. [CRG] (1951) The problem of serial order in behavior. In: Cerebral mechanisms in behavior, ed. L. A. Jeffress, pp. 112-36. New York: John Wiley 4 Sons. [DCM] Lee, D. N. (1980) Visuo-motor coordination in space-time. In: Tutorials in motor behavior, ed. C. Stelmach & J. Requin. Amsterdam: North Holland. [CRG, ESR] Leyhausen, P. (1965) Uber die Funktion der relativen Stimmungs-hierarchie (Dargestellt am Beispiel der phylogenetischen und ontogenetischen Entwicklung des Beutefangs von Raubtieren). Zeitschrift fur Tierpsychologie 22:412-94. [CRG] Liberman, A. M. (1980) An ethiological approach to language through the study of speech perception. In: Human ethology, ed. M. von Cranch, K. Foppa, W. Lepenies, 4 D. Ploog, pp. 682-704. Cambridge University Press. [JCF] Linsenmair, K. E. (1969) Anemotaktische Orientierung bei Tenebrioniden und Mistkafern (Insecta, Coleoptera). Zeitschrift fur vergleichende Physiologie 64:154-211. [CRC] (1970) Die Interaktion der paarigen antennalen Sinnesorgane bie der Wtadorientierung laufender Mist- und Schwarzkafer (Insecta, Coleoptera). Zeitschrift fur vergleichende Physiologie 70:247-77. [CRG] (1973) Die Windorientierung laufender Insekten. Fortschritte der Zoologie 21:59-80. [CRG] Lorenz, K. (1937) Uber die Bildung des Instinktbegriffs. Die Naturwissenschaften 25:289-300,307-18, 324-31. [CRG] (1950) The comparative method of studying innate behavior patterns. In: Symposia of the Society for Experimental Biology. No. 4: Physiological mechanisms in animal behavior. New York: Academic Press. [CRG] McCulloch, W. S. (1945) A heterarchy of values determined by the topology of nervous nets. Bulletin of Mathematical Biophysics 7:89-93. [MAA] MacDonnell, M. F. & Flynn, J. P. (1966) Sensory control of hypothalamic attack. Animal Behaviour 14:399-405. [CRG] McFarland, D. J. (1974) Motivational control systems analysis. New York: Academic Press. [WT] MacKay, D. G. & Bowman, R. W. (1969) On producing the meaning in sentences. American Journal of Psychology 82:23-39. [DGM] Mackintosh, N. J. 4 Dickinson, A. (1979) Instrumental (Type II) conditioning. In: Mechanisms of learning and motivation, ed. A. Dickinson 4 R. A. Boakes, pp. 143-69. Hillsdale, N.J.: Lawrence Erlbaum Associates. [NJM] Maier, N. R. F. (1929) Reasoning in white rats. Comparative Psychology Monographs 6, no. 29. [CRG, ESR] (1949) Frustration: The study of behavior without a goal. New York: McGraw-Hill. [DSO] Merton, P. A. (1973) How we control the contraction of our muscles. Scientific American 288:30-37. [CRG] Miller, S. & Konorski, J. (1928) Sur une forme particuliere des reflexes conditionnels. Comptes Rendus des Seances de la Societe de Biologie. 99, 1155-57. [NJM] Milner, P. M. (1961) The application of physiology to learning theory. In: Current trends in psychological theory, ed. R. A. Patlon, pp. 111-33. Pittsburgh: University of Pittsburgh Press. [PMM] Minsky, M. 4 Papert, S. (1972) Artificial intelligence. AI Memo 252, M.I.T. [MAA] Morgan, C. T. (1943) Physiological psychology. New York: McGraw-Hill. [CRG] Nashner, L. M.; Woollacott, M.; & Tuma, G. (1979) Organization of rapid responses to postural and locomotor-like perturbations in standing man. Experimental Brain Research 36:463-76. [CRG] Newell, A. (1980) Physical symbol systems. Cognitive Science 4:135-83. [AN] Norman, M. F. (1981) Lectures on linear systems theory. Journal of Mathematical Psychology 23:1-89. [CRG] O'Keefe, J. 4 Nadel, L. (1978) The hippocampus as a cognitive map. Oxford: Clarendon Press. [CRG] Olton, D. S. (1979) Mazes, maps, and memory. American Psychologist 34:58896. [DSO] Olton, D. S. 4 Samuelson, R. J. (1976) Remembrance of places passed: Spatial memory in rats. Journal of Experimental Psychology: Animal Behavior Processes 2:97-116. [CRG] Pantin, D. F. A. (1965) Learning, world-models and pre-adaptation. Animal Behavior Supplements 1:1-8. [ESR] Pearson, K. G. (1976) The control of walking. Scientific American 235:72-86. [CRG] Pearson, K. G.; Heitler, W. J.; 4 Stevens, J. D. (1980) Triggering of locust jump by multimodal inhibitory interneurons. Journal of Neurophysiology 43:257-78. [RWD] Pew, R. W. (1974) Human perceptual motor performance. In: Human information processing: Tutorials in performance and cognition, ed. B. H. Kantowitz. Hillsdale, N.J.: Lawrence Erlbaum Associates. [CRC] Pinker, S. 4 Kosslyn, S. M. (1978) The representation and manipulation of three-dimensional space in mental images. Journal of Mental Imagery 2:69-84. [CRC]

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4

649

References/Gallistel: Organization of action

Premack, D. (1962) Reversibility of the reinforcement relation. Science 136:255-57. [CRG] Preyer, W. (1885) Specielle Physiologie des Embryo. Leipzig: Grieben's Verlag. [RRP] Provine, R. R. (1976) Development of function in nerve nets. In: Simpler networks and behavior, ed. J. C. Fentress. Sunderland, Mass.: Sinauer Associates. [RRP] (1979) "Wing-flapping" develops in wingless chicks. Behavioral and Neural Biology 27:233-37. [RRP] (in press) Wing-flapping develops in chickens made flightless by feather mutations. Developmental Psychobiology. [RRP] Reed, E. S. (1980) The corporeal ideas hypothesis and the origin of scientific psychology. Doctoral dissertation, Boston University. Dissertation Abstracts International 41, 2152A (University Microfilms, no. 8024216). [ESR] Riggs, D. S. (1970) Control theory and physiological feedback mechanisms. Baltimore: Williams & Wilkins. [CRG] Robinson, D. A. (1973) Models of the saccadic eye movement control system. Kybemetik 14:71-83. [PD] Roitblat, H. L. (1980) Codes and coding processes in pigeon short-term memory. Animal Learning and Behavior 8:341-51. [HLR] (1981) The meaning of representation in animal memory. Submitted for publication. [HLR] Rose, S. P. R. (1981) From causations to translations: What biochemists can contribute to the study of behavior. In: Perspectives in ethology, volume 4, ed. P. P. G. Bateson It P. H. Klopfer, pp. 157-77. New York: Plenum Press. [JCF] Sacerdoti, E. D. (1974) Planning in a hierarchy of abstraction spaces. Artificial Intelligence 5:115-35. [MAA] Saltzman, E. (1979) Levels of sensorimotor representation. Journal of Mathematical Psychology 20:91-163. [AN] Salvini-Plawen, L. & Mayr, E. (1977) On the evolution of photoreceptors. Evolutionary Biology 10:207-63. [ESR] Schaffer, L. H. (1980) Analyzing piano performance: A study of concert pianists. In: Tutorials on motor behavior, ed. G. E. Stelmach & ]. Requin, pp. 443-55. Amsterdam: North-Holland Publishing Co. [MAA] Schmidt, R. A. & McGown, C. (1980) Terminal accuracy of unexpectedly loaded rapid movements: Evidence for a mass-spring mechanism in programming. Journal of Motor Behavior 12:149-61. [CRG] Selverston, A. I. (1980) Are central pattern generators understandable? Behavioral and Brain Sciences 3:535-71. [JCF, SG] Sevenster, P. (1973) Incompatibility of response and reward. In: Cons/rainls on learning, ed. R. A. Hinde & J. Stevenson-Hinde. New York: Academic Press. [CRG] Shepard, R. N. (1975) Form formation and transformation of internal representations. In: Information processing and cognition, ed. R. Solso. Hillsdale, N.J.: Erlbaum. [HRL] Sherrington, C. S. (1947) The integrative action of the nervous system. New Haven: Yale University Press. (First edition 1906) [CRG] Stark, L. (1959) Stability, oscillations and noise in the human pupil servo. Proceedings of the Institute of Radio Engineers 47:1925-39. [ERL] Stein, B. E.; Clamann, H. P.; & Goldberg, S. J. (1980) Superior colliculus: Control of eye movements in neonatal kittens. Science 210:78-80. [RRP] Stein, P. S. G. & Grossman, M. L. (1980) Central program for scratch reflex in turtle. Journal of Comparative Physiology A 140:287-94. [JCF] Stein, R. B. (1974) Peripheral control of movement. Physiological Reviews 54:215-44. [CRG] Stellar, E. (1960) Drive and motivation. In: Handbook of physiology, ed. J. Field & V. E. Hall. Section 1, vol. 3: Neurophysiology, ed. H. W. Magoun. Washington, D.C.: American Physiological Society. [CRG] Stelmach, G. E. & Requin, J., eds. (1980) Tutorials in Motor Behavior, NorthHolland Publishing Co. [MAA, CRG] Studdert-Kennedy, M. (1976) Speech perception. In: Contemporary issues in experimental phonetics, ed. N. J. Lass. New York: Academic Press. [JCF] Teitelbaum, P. (1971) The encephalization of hunger. Progress in Physiological Psychology 4:319-50. [CRG] Thorpe, W. H. (1950) A note on detour behaviour with Ammophila pubescens Curt. Behaviour 2:257-64. [CRG] (1963) Learning and instinct in animals. Cambridge: Harvard University Press. [CRG] Timberlake, W. (1980) A molar equilibrium theory of learned performance. In: The psychology of learning and motivation, vol. 14, ed. G. H. Bower, pp. 1-58. New York: Academic Press. [WT] (1981) The functional organization of appetitive behavior: Behavior systems and learning. In: Advances in analysis of behaviour, vol. 3: Biological factors in learning, ed. M. D. Zeiler & P. Harzem. Chichester: Wiley. [WT] Tinbergen, N. (1951) The study of instinct. London: Oxford University Press. [CRG, JAH, WT] Tinbergen, N. & Kruyt, W. (1938) Uber die Orientierung des Bienenwolfes (Philanthus triangulum Fabr.). III. Die Bevorzugung bestimmter Wegmarken. Zeitschrift fur vergleichende Physiologie 25:292-334. [CRG) Tinkelpaugh, O. L. (1932) Multiple delayed reactions with chimpanzees and monkeys. Journal of Comparative Psychology 13:207-43. [CRG] Tolman, E. C ; Ritchie, B. F.; 4 Kalish, D. (1946) Studies in spatial learning. I. Orientation and the short-cut. Journal of Experimental Psychology 36:13-24. [RJ] Turvey, M. T. (1977) Preliminaries to a theory of action with reference to vision. In: Perceiving, acting, and knowing, ed. R. Shaw & J. Bransford. Hillsdale: N.J.: Lawrence Erlbaum Associates. [CRG, SG] Turvey, M. T.; Shaw, R. E.; Reed, E. S.; & Mace, W. M. (1981) Ecological laws for perceiving and acting: In reply to Fodor and Pylyshyn. Cognition, in press. [ESR] Twitchell, T. E. (1970) Reflex mechanisms and the development of prehension. In: Mechanisms of motor skill development, ed. K. Connally. New York: Academic Press. [CRG] Uttal, W. R. (1978) The psychobiology of mind. Hillsdale, N.J.: Lawrence Erlbaum Associates. [JCF] Weiss, P. (1941) Self-differentiation of the basic patterns of coordination. Comparative Psychology Monographs, vol. 17, no. 4. [CRG] Wickens, D. D. (1938) The transference of conditioned excitation and conditioned inhibition from one muscle group to the antagonistic muscle group. Journal of Experimental Psychology 22:101-23. [CRG, DGM] Weiss, P. (1941) Self-differentiation of the basic patterns of coordination. Comparative Psychology Monographs 17(4). [JAH, JMH] (1939) A study of voluntary and involuntary finger conditioning. Journal of Experimental Psychology 25:127-40. [CRC, HLR] Wilson, D. M. (1966) Insect walking. Annual Review of Entomology 11:10322. [CRG, JAH] Woolridge, M. W. (1975) A quantitative analysis of short-term rhythmical behaviour in rodents. Doctoral dissertation, Oxford University. LJCF] Wright, P. J. (1979) On a clear day you can see General Motors. New York: Avon. [RCB] Young, R. M. (1970) Mind, brain, and adaptation in the nineteenth century. Oxford: Clarendon Press. [CRG]

650

THE BEHAVIORAL AND BRAIN SCIENCES (1981), 4