Supervised Learning I:

Perceptrons and LMS

Instructor: Tai-Yue (Jason) Wang

Department of Industrial and Information Management
Institute of Information Management

1/120
 Two Fundamental Learning
Paradigms
 Non-associative
 an organism acquires the properties of a single
repetitive stimulus.
 Associative
 an organism acquires knowledge about the
relationship of either one stimulus to another, or one
stimulus to the organism’s own behavioral response to
that stimulus.

2/120
 Examples of Associative
Learning(1/2)
 Classical conditioning
 Association of an unconditioned stimulus (US) with a
conditioned stimulus (CS).
 CS’s such as a flash of light or a sound tone produce
weak responses.
 US’s such as food or a shock to the leg produce a
strong response.
 Repeated presentation of the CS followed by the US,
the CS begins to evoke the response of the US.
 Example: If a flash of light is always followed by
serving of meat to a dog, after a number of learning
trials the light itself begins to produce salivation. 3/120
 Examples of Associative
Learning(2/2)
 Operant conditioning
 Formation of a predictive relationship between a
stimulus and a response.
 Example:
Place a hungry rat in a cage which has a lever
on one of its walls. Measure the spontaneous
rate at which the rat presses the lever by virtue
of its random movements around the cage. If
the rat is promptly presented with food when
the lever is pressed, the spontaneous rate of
lever pressing increases!
4/120
 Reflexive and Declarative Learning
 Reflexive learning
 repetitive learning is involved and recall does not
involve any awareness or conscious evaluation.
 Declarative learning
 established by a single trial or experience and
involves conscious reflection and evaluation for its
recall.
 Constant repitition of declarative knowledge
often manifest itself in reflexive form. 5/120
 Important Aspects of Human
Memory(1/4)
Input
 Two distinct stages: stimulus

 short-term memory (STM) SHORT TERM MEMORY

(STM)

 long-term memory (LTM) Download Recall

Process
Memory
recalled
LONG TERM MEMORY
 Inputs to the brain are (LTM)

processed into STMs Recall process of these memories

which last at the most for a is distinct from the memories
themselves.
few minutes.

6/120
 Important Aspects of Human
Memory(2/4)
Input
 Information is stimulus

downloaded into SHORT TERM MEMORY

(STM)

LTMs for more Download Recall

Process
Memory
recalled

permanent storage: LONG TERM MEMORY

(LTM)

days, months, and

Recall process of these memories
years. is distinct from the memories
themselves.
 Capacity of LTMs is
very large.
7/120
Important Aspects of Human
Memory(3/4)
 Recall of recent memories is more easily
disrupted than that of older memories.
 Memories are dynamic and undergo
continual change and with time all
memories fade.
 STM results in
 physical changes in sensory receptors.
 simultaneous and cohesive reverberation of
neuron circuits.

8/120
 Important Aspects of Human
Memory(4/4)
 Long-term memory involves
 plastic changes in the brain which take the form of
strengthening or weakening of existing synapses
 the formation of new synapses.
 Learning mechanism distributes the memory
over different areas
 Makes robust to damage
 Permits the brain to work easily from partially
corrupted information.
 Reflexive and declarative memories may
actually involve different neuronal circuits. 9/120
 Learning Algorithms(1/2)
 Define an architecture-dependent procedure to
encode pattern information into weights
 Learning proceeds by modifying connection
strengths.
 Learning is data driven:
 A set of input–output patterns derived from a (possibly
unknown) probability distribution.
 Output pattern might specify a desired system response for a
given input pattern
 Learning involves approximating the unknown function as
described by the given data. 10/120
 Learning Algorithms(2/2)
 Learning is data driven:
 Alternatively, the data might comprise patterns that
naturally cluster into some number of unknown
classes
 Learning problem involves generating a suitable
classification of the samples.

11/120
 Supervised Learning(1/2)
 Data comprises a set of An example function described
by a set of noisy data points
discrete samples 20

T = { ( X k , Dk )}
5 4 3 2
3 x -1.2 x -12.27 x +3.288 x +7.182 x
Q 15

k =1 10

drawn from the pattern 5

space where each

f(x)
0

-5

sample relates an input -10

vector Xk ∈ Rn to an -15

-20

output vector Dk ∈ Rp.

-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x

12/120
 Supervised Learning(2/2)
 The set of samples
describe the behavior 20
An example function described
by a set of noisy data points
5 4 3 2

of an unknown
3 x -1.2 x -12.27 x +3.288 x +7.182 x
15

10

function f : Rn → Rp 5

which is to be
f(x)
0

-5

characterized. -10

-15

-20
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x

13/120
 The Supervised Learning Procedure
Error information fed back for network adaptation

Sk Error
Xk

Dx
Neural Network

 We want the system to generate an output Dk in

response to an input Xk, and we say that the system has
learnt the underlying map if a stimulus Xk’ close to Xk
elicits a response Sk’ which is sufficiently close to Dk.
14/120
The result is a continuous function estimate.
 Unsupervised Learning(1/2)
 Unsupervised learning provides the system with
an input Xk, and allow it to self-organize its
weights to generate internal prototypes of sample
vectors. Note: There is no teaching input involved
here.
 The system attempts to represent the entire data
set by employing a small number of prototypical
vectors—enough to allow the system to retain a
desired level of discrimination between samples.
15/120
 Unsupervised Learning(2/2)
 As new samples continuously buffer the system,
the prototypes will be in a state of constant flux.
 This kind of learning is often called adaptive
vector quantization

16/120
 Clustering and Classification(1/3)
 Given a set of data
samples {Xi}, Xi ∈ Rn, 2.5

Cluster 1

is it possible to identify
2

1.5
well defined “clusters”, 1

where each cluster 0.5

Cluster 2 Cluster centroids

defines a class of 0

vectors which are -0.5

similar in some broad -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5

sense? 17/120
 Clustering and Classification(2/3)
 Clusters help establish a
classification structure 2.5

Cluster 1

within a data set that has

2

1.5

no categories defined in 1

advance. 0.5
Cluster 2 Cluster centroids

 Classes are derived from 0

clusters by appropriate
-0.5

labeling.
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5

18/120
 Clustering and Classification(3/3)
 The goal of pattern classification
is to assign an input pattern to one
2.5
of a finite number of classes. Cluster 1
2
 Quantization vectors are called
codebook vectors. 1.5

Cluster 2 Cluster centroids

0.5

-0.5

-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5

19/120
Characteristics of Supervised and
Unsupervised Learning

20/120
 General Philosophy of Learning:
Principle of Minimal Disturbance
 Adapt to reduce the output error for the
current training pattern, with minimal
disturbance to responses already learned.

21/120
 Error Correction and Gradient
Descent Rules
 Error correction rules alter the weights of a
network using a linear error measure to reduce
the error in the output generated in response to
the present input pattern.
 Gradient rules alter the weights of a network
during each pattern presentation by employing
gradient information with the objective of
reducing the mean squared error (usually
averaged over all training patterns).
22/120
 Learning Objective for
TLNs(1/4)
+1

 Augmented Input and X1

W1
WO

Weight vectors
X k = ( x0 , x1k , , x )
W2
k T
n X k ∈ R n+1 X2
Σ S

Wk = ( w , w , , w )W

…
Wn
k T n+1
k
0
k
1 n k ∈R
Xn
 Objective: To design the
weights of a TLN to
correctly classify a given
set of patterns.
23/120
 Learning Objective for
TLNs(2/4)
+1
 Assumption: A training
WO
set of following form is X1
W1

given W2

T = { ( X k , Dk )}
Q
XK ∈R n +1
, d k ∈ { 0,1}
X2
Σ S
k =1

…
Wn

 Each pattern Xk is tagged

Xn
to one of two classes C0 or
C1 denoted by the desired
output dk being 0 or 1
respectively. 24/120
 Learning Objective for TLNs(3/4)
 Two classes identified by two possible signal
states of the TLN
 C0 by a signal S(yk) = 0, C1 by a signal S(yk) = 1.
 Given two sets of vectors X0 andX1 belonging
to classes C0 and C1 respectively the learning
procedure searches a solution weight vector
WS that correctly classifies the vectors into
their respective classes.

25/120
 Learning Objective for TLNs(4/4)
 Context: TLNs
 Find a weight vectorWS such that for all Xk ∈ X1,
S(yk) = 1; and for all Xk ∈ X0, S(yk) = 0.
 Positive inner products translate to a +1 signal
and negative inner products to a 0 signal
 Translates to saying that for all Xk ∈ X1, XkTWS >
0; and for all Xk ∈ X0, XkTWS < 0.

26/120
 Pattern Space(1/2)
Activation
 Points that satisfy XTWS =
0 define a separating
hyperplane in pattern
space.
 Two dimensional case:
 Pattern space points on one
side of this hyperplane
(with an orientation
indicated by the arrow)
yield positive inner
products with WS and thus
generate a +1 neuron
signal. 27/120
 Pattern Space(2/2)
Activation
 Two dimensional case:
 Pattern space points on
the other side of the
hyperplane generate a
negative inner product
with WS and consequently
a neuron signal equal to 0.
 Points in C0 and C1 are
thus correctly classified
by such a placement of
the hyperplane.
28/120
 A Different View: Weight Space
(1/2)
 Weight vector is a variable
vector.
 WTXk = 0 represents a
hyperplane in weight space
 Always passes through the
origin since W = 0 is a
trivial solution of WTXk = 0.

29/120
 A Different View: Weight Space
(2/2)
 Called the pattern
hyperplane of pattern Xk.
 Locus of all pointsW such
thatWTXk = 0.
 Divides the weight space
into two parts: one which
generates a positive inner
product WTXk > 0, and the
other a negative inner
product WTXk<0.

30/120
 Identifying a Solution Region from
Orientated Pattern Hyperplanes(1/2)
W2
X3 X2
 For each pattern Xk in
pattern space there is a X1
corresponding
hyperplane in weight
space. W1

 For every point in weight

space there is a X4

corresponding
hyperplane in pattern
Solution region
space.
31/120
 Identifying a Solution Region from
Orientated Pattern Hyperplanes(2/2)
W2
X3 X2
 A solution region in
weight space with X1

four pattern
hyperplanes W1
 χ
1 = {X1,X2}
 χ
0 = {X3,X4}
X4

Solution region

32/120
 Requirements of the Learning
Procedure
 Linear separability guarantees the existence of a
solution region.
 Points to be kept in mind in the design of an
automated weight update procedure :
 It must consider each pattern in turn to assess the correctness of the
present classification.
 It must subsequently adjust the weight vector to eliminate a
classification error, if any.
 Since the set of all solution vectors forms a convex cone, the
weight update procedure should terminate as soon as it penetrates
the boundary of this cone (solution region). 33/120
 Design in Weight Space(1/3)
 Assume: Xk ∈ X1 and Xk
Wk+1
WkTXk as erroneously
non-positive.
 For correct
classification, shift the WkT Xk>0
weight vector to some Wk
position Wk+1 where the
inner product is WT Xk<0

positive.

34/120
 Design in Weight Space(2/3)
Xk
 The smallest Wk+1

perturbation in Wk that
produces the desired
change is, the
perpendicular distance WkT Xk>0
Wk
from Wk onto the
pattern hyperplane. WT Xk<0

35/120
 Design in Weight Space(3/3)
 In weight space, the Xk
Wk+1
direction
perpendicular to the
pattern hyperplane is
none other than that WkT Xk>0
of Xk itself. Wk

WT Xk<0

36/120
 Simple Weight Change Rule:
Perceptron Learning Law

 If Xk ∈ X1 and WkTXk < 0 add a fraction of the pattern to

the weight Wk if one wishes the inner product WkTXk to
increase.
 Alternatively, if Xk ∈ X0, and WkTXk is erroneously non-
negative we will subtract a fraction of the pattern from
the weight Wk in order to reduce this inner product.
37/120
 Weight Space Trajectory
W2
X3 X2
 The weight space
trajectory corresponding X1
to the sequential
presentation of four
patterns with pattern W1

hyperplanes as indicated:
 χ 1 = {X1,X2} and X4
χ 0 = {X3,X4}

Solution region

38/120
 Linear Containment
 Consider the set X0’ in which each element X0 is
negated.
 Given a weight vectorWk, for any Xk ∈ X1 ∪ X0’, XkT
Wk > 0 implies correct classification and XkTWk < 0
implies incorrect classification.
 X‘ = X1 ∪ X0’ is called the adjusted training set.
 Assumption of linear separability guarantees the
existence of a solution weight vector WS, such that
XkTWS > 0 ∀ Xk ∈ X
 We say X’ is a linearly contained set. 39/120
 Recast of Perceptron Learning with
Linearly Contained Data

 Since Xk ∈ X’, a misclassification of Xk will add

η kXk to Wk.
 ForXk ∈ X0‘, Xk actually represents the negative of
the original vector.
 Therefore addition of η kXk to Wk actually amounts
to subtraction of the original vector from Wk.
40/120
Perceptron Algorithm:
Operational Summary

41/120
 Perceptron Convergence
Theorem
 Given: A linearly contained training set X’ and
any initial weight vectorW1.
 Let SW be the weight vector sequence generated in
response to presentation of a training sequence SX
upon application of Perceptron learning law. Then
for some finite index k0 we have: Wk0 = Wk0 +1 =
Wk0 +2 = ・ ・ ・ = WS as a solution vector.
 See the text for detailed proofs.
42/120
 Hand-worked Example
+1

X1
WO

W1

W2

X2

Binary threshold neuron

43/120
Classroom Exercise(1/2)

44/120
Classroom Exercise(2/2)

45/120
MATLAB Simulation
(a) Hyperplane movement depicted during Perceptron Learning
3
k=15 2
R
2.5

2
k=35
1.5
k=25
2
x

1
(0,1) k=5 (1,1)

0.5

0
(0,0) (1,0)

-0.5
-0.2 0 0.2 0.4 0.6 0.8 1 1.2
x1

(b) Weight space trajectories: W 0 = (0,0,0), W S =(-4 3 2)

3
0
2
w2

1 -1

0 -2
3
2.5 -3
2
1.5
1
0.5 -4
0
w0
w
1

46/120
 Perceptron Learning and Non-
separable Sets
 Theorem:
 Given a finite set of training patterns X, there
exists a number M such that if we run the
Perceptron learning algorithm beginning with
any initial set of weights,W1, then any weight
vector Wk produced in the course of the
algorithm will satisfyWk ≤ W1 +M

47/120
 Two Corollaries(1/2)
 If, in a finite set of training patterns X, each
pattern Xk has integer (or rational) components
xik, then the Perceptron learning algorithm will
visit a finite set of distinct weight vectors Wk.

48/120
 Two Corollaries(2/2)
 For a finite set of training patterns X, with
individual patterns Xk having integer (or rational)
components xik the Perceptron learning algorithm
will, in finite time, produce a weight vector that
correctly classifies all training patterns iff X is
linearly separable, or leave and re-visit a specific
weight vector iff X is linearly non-separable.

49/120
 Handling Linearly Non-separable
Sets: The Pocket Algorithm(1/2)
 Philosophy: Incorporate positive reinforcement
in a way to reward weights that yield a low error
solution.
 Pocket algorithm works by remembering the
weight vector that yields the largest number of
correct classifications on a consecutive run.

50/120
 Handling Linearly Non-separable
Sets: The Pocket Algorithm(2/2)
 This weight vector is kept in the “pocket”, and
we denote it as Wpocket .
 While updating the weights in accordance with
Perceptron learning, if a weight vector is
discovered that has a longer run of consecutively
correct classifications than the one in the pocket,
it replaces the weight vector in the pocket.

51/120
Pocket Algorithm:
Operational Summary(1/2)

52/120
Pocket Algorithm:
Operational Summary(2/2)

53/120
Pocket Convergence Theorem
 Given a finite set of training examples, X,
and a probabilityp < 1, there exists an
integer k0 such that after any k > k0
iterations of the pocket algorithm, the
probability that the pocket weight
vectorWpocket is optimal exceeds p.

54/120
 Linear Neurons and Linear Error
 Consider a training set of the form T = {Xk, dk},
Xk ∈ Rn+1 , dk ∈ R.
 To allow the desired output to vary smoothly or
continuously over some interval consider a linear
signal function: sk = yk = XkTWk
 The linear error ek due to a presented training pair
(Xk, dk), is the difference between the desired
output dk andthe neuronal signal sk: ek = dk − sk
= dk − XkTWk 55/120
 Operational Details of α –LMS
+1

 α –LMS error correction X1k

WO
is proportional to the error W k
1

itself Xk
W2k
Sk=XkTWk
2
 Each iteration reduces the

…
error by a factor of η. Wnk
Xn k

 η controls the stability and

speed of convergence.
 Stability ensured
if 0 < η < 2.
56/120
 α –LMS Works with Normalized
Training Patterns
x3

xk

Wk+1

Wk

Wk
x2

x1

57/120
α –LMS: Operational Summary

58/120
 MATLAB Simulation Example
 Synthetic data set 10

shown in the figure is

generated by artificially
5

f = 3x(x-1)(x-1.9)(x+0.7)(x+1.8)± ε
scattering points around 0

a straight line: y = 0.5x -5

+ 0.333 and generate a -1 0

scatter of 200 points in -1 5

a ±0.1 interval in the y -2 0

direction.
-2 -1 0 1 2
x

59/120
 MATLAB Simulation Example
 This is achieved by first 10

generating a random
scatter in the interval
5

f = 3x(x-1)(x-1.9)(x+0.7)(x+1.8)± ε
[0,1]. 0

Then stretching it to the

-5

interval [−1, 1], and -1 0

finally scaling it to ±0.1 -1 5

-2 0
-2 -1 0 1 2
x

60/120
Computer Simulation of α -LMS
Algorithm
2

1 .8

Ite ra tio n 1
1 .6

1 .4

1 .2
y

0 .8

0 .6

Ite ra tio n 5 0 (M S E = 0 .0 4 )
0 .4
Ite ra tio n 1 0

0 .2

0 .2 0 .4 0 .6 0 .8 1 1 .2 1 .4 1 .6 1 .8 2
x

61/120
 A Stochastic Setting
 Assumption that the training set T is well defined
in advance is incorrect when the setting is
stochastic.
 In such a situation, instead of deterministic
patterns, we have a sequence of samples {(Xk, dk)}
assumed to be drawn from a statistically
stationary population or process.
 For adjustment of the neuron weights in response
to some pattern-dependent error measure, the error
computation has to be based on the expectation of
the error over the ensemble. 62/120
 Definition of Mean Squared Error
(MSE)(1/2)
 We introduce the square error on a pattern Xk as
ε k = ( d k − X k Wk ) = ek2
1 T 2 1
(1)
2 2
= ( d k − 2d k X kT Wk + WkT X k X kT Wk )
1 2
(2)
2
 Assumption: The weights are held fixed at Wk
while computing the expectation.

63/120
 Definition of Mean Squared Error
(MSE)(2/2)
 The mean-squared error can now be computed by
taking the expectation on both sides of (2):
1
[ ] [ ] 1 T
[ ]
ε = E [ε k ] = E d k − E d k X k Wk + Wk E X k X kT Wk
2
2 T

2
(3)

64/120
Our problem…
 To find optimal weight vector that
minimizes the mean-square error.

65/120
 Cross Correlations(1/2)
 For convenience of expression we define the
pattern vector P as the cross-correlation between
the desired scalar output, dk, and the input vector,
Xk
[ ] [
P T ∆ E d k X kT = E ( d k , d k x1k ,, d k xnk ) ] (4)

 and the input correlation

1 x
matrix,
 x 
R, as
k k
1 n
 k 
[ ]
k k k k
x x x  x x 
R ∆ E X k X kT = E 1 1 1 1 n
(5)
    
 
66/120
 x nk k k
xn x1 x n xn 
k k
Cross Correlations (2/2)
 Using Eqns. (4)-(5), we rewrite the MSE
expression of Eqn. (3) succinctly as

[ ]
1
2
2 T 1 T
ε = E [ε k ] = E d k − P Wk + Wk RWk
2
(6)

 Note that since the MSE is a quadratic

function of the weights it represents a bowl
shaped surface in the (n+1) x 1 weight—MSE
cross space.
67/120
 Finding the Minimum Error(1/2)
 First compute the gradient by straightforward
differentiation which is a linear function of weights

T
 ∂ε ∂ε 
∇ε =  , ,  = − P + RW (7)
 To find ∂the
w0 optimal
∂wn set of weights, , simply set
Ŵ
which yields
∇ε = 0

RWˆ = P (8) 68/120

 Finding the Minimum Error (2/2)
 This system of equations (8) is called the Weiner-
Hopf system and its solution is the Weiner
solution or the Weiner filter
Wˆ = R −1 P (9)
 is the point in weight space that represents the
minimum mean-square error ε min

69/120
 Computing the Optimal
Filter(1/2)
 First compute R and P . Substituting from Eqn.
−1

(9) into Eqn. (6) yields:

ε min
1
2
[ ]
2 1 ˆT ˆ
= E d k + W RW − P TWˆ
2
(10)

1
2
[ ]
= E d k + ( R P ) R ( R −1P ) − P T R −1 P
2 1 −1 T
2
(11)

1
[ ] 1
= E d k2 − P TWˆ
2 2
(12)

70/120
 Computing the Optimal
Filter(2/2)
 For the treatment of weight update procedures
we reformulate the expression for mean-square
error in terms of the deviation
V = W − Wˆ , of the
weight vector from the Weiner solution.

71/120
 Computing R(1/2)
 Substituting ˆ Eqn. (6)
into
W =V +W
[ ]
2
k
1
2
( ) (
T
) (
ε = E d + V + W R V + Wˆ − P T V + Wˆ
ˆ ) (13)

1  T 
[ ]
= E d k + V RV + V 
2

2 
T
RWˆ + Wˆ RV + Wˆ RWˆ  − P TV − P TWˆ
  
T T

ˆ T T −1 T
2W RV = 2 P R RV = 2 P V 
(14)
1 T
= ε min + V RV (15)
2
(
1
2
) (
T
ˆ T
) (
= ε min + W − W R W − W R W − Wˆ
ˆ ) (16)

72/120
 Computing R(2/2)
 Note that since the mean-square error ε is
non-negative, we must have V RV ≥ 0 . This
T

implies that R is positive semi-definite.

Usually however, R is positive definite.

73/120
Diagonalization of R
 Assume that R has distinct eigenvalues λi . Then we
can construct a matrix Q whose columns are
corresponding eigenvectors ηi of R.
Q = (η0 η1  η n ) (17)
 R can be diagonalized using an orthogonal similarity
transformation as follows. Having constructed Q, and
knowing that:  λ0 0  0 
 
We have  0 λ1  0 
RQ = (η0 η1  η n )  (18)
  
 
0  λ 
 n

Q −1 RQ = D = diag{ λ0 , λ1 , λ2 ,, λn } (19) 74/120

 Some Observations
 It is usual to choose the eigenvectors of R to be
orthonormal QQ T = I and Q −1 = Q.T
 Then R = QDQT = QDQ −1
 From Eqn. (15) we know that the shape of ε is a bowl-
shaped paraboloid with a minimum at the Weiner
solution (V=0, the origin of V-space).
 Slices of ε parallel to the W space yield elliptic
constant error contours which define the weights in
weight space that the specific value of the square-error
(say ε c) at which the slice is made:
1 T
V RV = εc − ε min = constant (20)
2 75/120
 Eigenvectors of R(1/2)
 Also note from Eqn. (15) that we can compute the
MSE gradient in V-space as,
∇ε = RV (21)
which defines a family of vectors in V-space.

 Exactly n+1 of these pass through the origin of V-

space and these are the principal axes
V , of the ellipse.
'

76/120
 Eigenvectors of R(2/2)
 However, vectors passing through the origin
must take the λV . Therefore, for the principal
axes V ' ,
RV ' = λV ' (22)
 Clearly, V ' is an eigenvector of R.

77/120
 Steepest Descent Search with Exact
Gradient Information
 Steepest descent search uses
exact gradient information
available from the mean-
square error surface to direct
the search in weight space.
 The figure shows a
projection of the square-error
function on the plane.

78/120
ε − wik
 Steepest Descent Procedure
Summary(1/2)
 Provide an appropriate weight increment to wik to
push the error towards the minimum which
occurs at ŵi .
 Perturb the weight in a direction that depends on
which side of the optimal weight ŵ the current
i

weight value w lies.

k
i

 If the weight component w lies to the left of ŵ ,

k
i i

say at w , where the error gradient is negative (as

k
i1

indicated by the tangent) we need to increase w . k

i

79/120
 Steepest Descent Procedure
Summary(2/2)
 If w is on the right of , say at
k
i
ŵi where the
wik2
error gradient is positive, we need to decrease
w k
i .
 This rule is summarized in the following
statement:
∂ε
If
∂wik
> 0,( wk
i > ˆ
w i ) , decrease
w k
i

∂ε
If
∂wik
< 0,( wk
i < ˆ
w i ) , increase
w k
i

80/120
 Weight Update Procedure(1/3)
 It follows logically therefore, that the weight
component should be updated in proportion
with the negative of the gradient:
 ∂ε 
wk +1
i = w + η  −
k
i  i = 0,1, , n ( 27 )
 ∂wi
k


81/120
 Weight Update Procedure(2/3)
 Vectorially we may write
Wk +1 = Wk + η ( − ∇ε ) (28)

where we now re-introduced the iteration time

dependence into the weight components:
T
 ∂ε ∂ε 
∇ε =  k , , k  (29)
 ∂w0 ∂wn 

82/120
 Weight Update Procedure(3/3)
 Equation (28) is the steepest descent update
procedure. Note that steepest descent uses
exact gradient information at each step to
decide weight changes.

83/120
Convergence of Steepest Descent –
1(1/2)
 Question: What can one say about the
stability of the algorithm? Does it converge
for all values of η ?
 To answer this question consider the
following series of subtitutions and
transformations. From Eqns. (28) and (21)
Wk +1 = Wk + η ( − ∇ε ) (30)
= Wk − ηRVk (31)
(
= Wk + ηR Wˆ − Wk ) (32)
= (1 − ηR )Wk + ηRWˆ (33)
84/120
Convergence of Steepest Descent –
1(2/2)

 Transforming Eqn. (33) into V-space (the

principal coordinate system) by substitution
of Vk + Wˆ for Wk yields:

Vk +1 = ( I − ηR )Vk (34)

85/120
 Steepest Descent Convergence –
2(1/2)
 Rotation to the principal axes of the elliptic
contours can be effected by using V = QV :
'

QVk'+1 = ( I − ηR ) QVk' (35)

or
Vk'+1 = Q −1 ( I − ηR ) QVk' (36)
( )
= I − ηD Vk'
(37)

86/120
 Steepest Descent Convergence –
2(2/2)
 where D is the diagonal eigenvalue matrix.
Recursive
or application of
Eqn. (37) yields:
Vk = ( I − ηD ) V0
' k '
(38)
 It follows from this that for stability and
convergence of the algorithm:
lim( I − ηD ) = 0
k
(39)
k →∞

87/120
 Steepest Descent Convergence –
3(1/2)
This requires that
lim(1 − η λmax ) = 0
k
(40)
k →∞

or
2
0 <η < (41)
λmax

λmax being the largest eigenvalue of R.

88/120
 Steepest Descent Convergence –
3(2/2)
If this condition is satisfied then we have

k →∞ k →∞
( )
lim Vk' = lim Q −1 Wk − Wˆ = 0 (42)

or lim Wk = Wˆ (43)
k →∞

 Steepest descent is guaranteed to converge to the

Weiner solution as long as the learning rate is
maintained within the limits defined by Eqn. (41).

89/120
 Computer Simulation
Example(1/2)
 This simulation example employs the fifth
order function data scatter with the data
shifted in the y direction by 0.5.
Consequently, the values of R,P and the
Weiner solution are respectively:
1 0.500   0.8386 
R =   , P =   (44)
 0.500 1.61   − 1.4625 
ˆ  1.5303 
W =   (45)
 − 1.3834 
90/120
 Computer Simulation
Example(2/2)
 Exact gradient information is available
since the correlation matrix R and the cross-
correlation matrix P are known.
 The weights are updated using the equation:
(
Wk +1 =W k+2ηR Wˆ − Wk ) (46)

91/120
 MATLAB Simulation
Example(1/3)
eta = .01; % Set learning rate
R=zeros(2,2); % Initialize correlation matrix
X = [ones(1,max_points);x]; % Augment input vectors

P = (sum([d.*X(1,:); d.*X(2,:)],2))/max_points; % Cross-correlations

D = (sum(d.^2))/max_points; %squares; % target expectations

for k =1:max_points
R = R + X(:,k)*X(:,k)'; % Compute R
end

R = R/max_points;
weiner=inv(R)*P; % Compute the Weiner solution
errormin = D - P'*inv(R)*P; % Find the minimum error

92/120
MATLAB Simulation Example (2/3)
shift1 = linspace(-12,12, 21); % Generate a weight space matrix
shift2 = linspace(-9,9, 21);
for i = 1:21 % Compute a weight matrix about
shiftwts(1,i) = weiner(1)+shift1(i); % Weiner solution
shiftwts(2,i) = weiner(2)+shift2(i);
end

for i=1:21 % Compute the error matrix

for j = 1:21 % to plot the error contours
error(i,j) = sum((d - (shiftwts(1,i) + x.*shiftwts(2,j))).^2);
end
end
error = error/max_points;

figure; hold on; % Plot the error contours

plot(weiner(1),weiner(2),'*k') % Labelling statements no shown
93/120
MATLAB Simulation Example (3/3)

w = 10*(2*rand(2,1)-1); % Randomize weights

w0 = w; % Remember the initial weights

for loop = 1:500 % Perform descent for 500 iters

w = w + eta*(-2*(R*w-P));
wts1(loop)=w(1); wts2(loop)=w(2);
End

% Set up weights for plotting

wts1=[w0(1) wts1]; wts2=[w0(2) wts2];

plot(wts1,wts2,'r') % Plot the weight trajectory

94/120
 Smooth Trajectory towards the
Weiner Solution
 Steepest descent
uses exact 8

6
160 W = (-3.9, 6.27) T
0
194 229 264

gradient
160

information to 2
55.3

search the 1
w
0
Weiner solution
(1.53, -1.38)
90.1
-2

Weiner solution -4 194

in weight space. -6
229
264
160

-8

368 299 125

-10 333 160
-10 -5 0 5 10
w
0

95/120
µ -LMS: Approximate Gradient
Descent(1/2)
 The problem with steepest descent is that
true gradient information is only available
in situations where the data set is
completely specified in advance.

 It is then possible to compute R and P

exactly, and thus the true gradient at
iteration
k : ∇ε = RWk − P
96/120
 µ -LMS: Approximate Gradient
Descent(2/2)
 However, when the data set comprises a
random stream of patterns (drawn from a
stationary distribution), R and P cannot be
computed accurately. To find a correct
approximation one might have to examine
the data stream for a reasonably large period
of time and keep averaging out.

 How long should we examine the stream to

get reliable estimates of R and P ?
97/120
 Definition
 The μ-LMS algorithm is convergent in the
mean if the average of the weight vector Wk
approaches the optimal solution Wopt as the
number of iterations k, approaches infinity:
 E[Wk] → Wopt as k → ∞

98/120
 µ -LMS employs ε k for
ε = Ε [ε k](1/2)
 The gradient computation modifies to:
T T
~  ∂ε ∂ε   ∂e ∂e 
∇ε k =  kk , kk  = ek  kk , , kk  = −ek X k (47)
 ∂w0 ∂wn   ∂w0 ∂wn 

 where , and since we are

ek = ( d k − sk ) = X k Wtherefore
T
dealing with linear neurons.skNote k that the
recursive update equation then becomes

~
Wk +1 = Wk + η (−∇ε k ) = Wk + η ( d k − sk ) X k (48)
= Wk + ηek X k (49)
99/120
 µ -LMS employs ε k for
ε = Ε [ε k](2/2)
~
 What value does the long term average of ∇ε k
converge to? Taking the Expectation of
both sides of Eqn. (47):
E [∇ε k ] = − E [ ek X k ] = − E [ d k X k − X k X kT W ]
~
(50)
= RW − P (51)
= ∇ε (52)

100/120
Observations(1/8)
 Since the long term average of approaches
~
∇ε k ,
~
we can safely ∇ε use as an unbiased
∇ε k estimate.
∇ε µ -LMS work!
That’s what makes

 Since approaches in the long run, one

could keep
~ collecting for a sufficiently large
number ∇ofε k iterations (while
∇ε keeping the weights
~
fixed), and then make a weightXk ∇ εk
change
collectively for all those iterations together.

101/120
Observations(2/8)
 If the data set is finite (deterministic), then
one can compute ∇ε accurately by first
collecting the different ∇~ ε gradients over
k

all training patterns

X
k for the same set of
weights. This accurate measure of the
gradient could then be used to change the
weights. In this situation µ -LMS is
identical to the steepest descent algorithm.

102/120
 Observations(3/8)
~
 Even if the data set is deterministic, we still use ∇ε k
to update the weights. After all if the data set becomes
large, collection of all the gradients becomes
expensive in terms of storage. Much easier to just go
ahead and use
 Be clear about the approximation made: we are
estimating the true gradient (which should be
computed from ) by a gradient computed from
E [ε ]
the instantaneous sample
k error . Although this may
seem to be a rather drastic approximation, ε k it works.

103/120
Observations(4/8)
 In the deterministic case we can justify this
as follows: if the learning rate η , is kept
small, the weight change in each iteration
will be small and consequently the weight
vector W will remain “somewhat constant”
over Q iterations where Q is the number of
patterns in the training set.

104/120
 Observations(5/8)
 Of course this is provided that Q is a small
number! To see this, observe the total weight
change ∇W , over Q iterations from the k th
iteration:
Q −1
∂ε k +i
∇W = − ∑ (53)
i =0 ∂Wk +i
 1 Q −1 ∂ε k +i 
= −Q ∑  (54)
 Q i =0 ∂Wk 
∂  1 Q −1 
= −Q  ∑ ε k +i  (55)
∂Wk  Q i =0 
∂ε
= −Q (56)
∂Wk 105/120
 Observations(6/8)
 Where ε denotes the mean-square error. Thus
the weight updates follow the true gradient on
average.

106/120
 Observations(7/8)
 Observe that steepest descent search is guaranteed to
search the Weiner solution provided the learning rate
condition (41) is satisfied.

 Now since is an unbiased estimate of one can

expect that µ -LMS
~ too will search out the Weiner
∇εto
∇ε kit does—but not smoothly. This is
solution. Indeed
be expected since we are only using an estimate of the
true gradient for immediate use.

107/120
 Observations(8/8)
 Although α -LMS and µ - LMS are similar
algorithms, α -LMS works on the normalizing
training set. What this simply means is that α -
LMS also uses gradient information, and will
eventually search out the Weiner solution-of the
normalized training set. However, in one case
the two algorithms are identical: the case when
input vectors are bipolar. (Why?)

108/120
µ -LMS Algorithm:
Convergence in the Mean (1)(1/2)
 Definition 0.1 The µ -LMS algorithm is convergent in
the mean if the average of the weight vector Wk
approaches the optimal solution Ŵas the number of
iterations k, approaches infinity:
E [Wk ] → Wˆ as k → ∞ (57)

 Definition 0.2 The µ -LMS algorithm is convergent in

the mean square if the average of the squared error ε k
approaches a constant as the Number of iterations, k,
approaches infinity:
E [ ε k ] → constant as k → ∞ (58)
109/120
 µ -LMS Algorithm:
Convergence in the Mean (1)(2/2)
 Convergence in the mean square is a
stronger criterion than convergence in the
mean. In this section we discuss
convergence in the mean and merely state
the result for convergence in the mean
square.

110/120
 µ -LMS Algorithm:
Convergence in the Mean (2)(1/2)
 Consider the µ -LMS weight update equation:
Wk +1 = Wk + η ( d k − sk ) X k (59)
= Wk + η ( d k − X kT Wk ) X k (60)
= Wk + ηd k X k − ηX kT Wk X k (61)
= Wk - ηX k X kT Wk + ηd k X k (62)
= ( I - ηX k X kT ) Wk + ηd k X k (63)

111/120
 µ -LMS Algorithm:
Convergence in the Mean (2)(2/2)
 Taking the expectation of both sides of Eqn.
(63) yields,
( [
E [Wk +1 ] = I − ηE X k X kT ] ) E[W ] + ηE[ d X ]
k k k (64)
= ( I − ηR ) E [Wk ] + ηP (65)

where P and R are as already defined

112/120
 µ -LMS Algorithm:
Convergence in the Mean (3)(1/2)
 Appropriate substitution yields:
( )
E [Wk +1 ] = I − ηQDQ T E [Wk ] + ηQDQ T Wˆ
= ( QQ )
− ηQDQ T E [Wk ] + ηQDQ T Wˆ
T
(66)
= Q( I − ηD ) Q T E [Wk ] + ηQDQ T Wˆ

 Pre-multiplication throughout by QT results in:

Q T E [Wk +1 ] = ( I − ηD ) Q T E [Wk ] + ηDQ T Wˆ (67)

113/120
µ -LMS Algorithm:
Convergence in the Mean (3)(2/2)
 And subtraction of Q TWˆ from both sides
gives:
( )
Q T E [Wk +1 ] − Wˆ = ( I − ηD ) Q T E [Wk ] − ( I − ηD ) Q T Wˆ (68)
(
= ( I − ηD ) Q T E [W ] − Wˆ
k ) (69)

 We will re-write Eqn. (69) in familiar

terms:
~ ~
Q T V = ( I − ηD ) Q T V (70)
k +1 k

114/120
 µ -LMS Algorithm:
Convergence in the Mean (4)(1/2)
And
~' ~'
( )
Vk +1 = I − ηD Vk (71)

where ~
and V~ = Q V~ . Eqn. (71)
Vk = E [Wk ] − Wˆ k
' T
k

represents a set of n+1Decoupled difference

equations :
v i = (1 − η λi ) v i
~ ' k +1 ~ 'k
i = 0,1, , n. (72)

115/120
 µ -LMS Algorithm:
Convergence in the Mean (4)(2/2)
Recursive application of Eqn. (72) yields,
v i = (1 − η λi ) v i
~ 'k k ~ '0
i = 0,1,  , n (73)

~ 'ik → 0
 To ensure convergence in the mean,
as v
k→∞ since this condition
requires that the
E [W ] deviation
k Ŵ of from
should tend to 0.

116/120
 µ -LMS Algorithm:
Convergence in the Mean (5)(1/2)
 Therefore from Eqn. (73):
| 1 − η λi | < 1 i = 0,1, , n. (74)
 If this condition is satisfied for the largest eigenvalue
then it will λbe satisfied for all other eigenvalues. We
max
therefore conclude that if
2
0 <η < (75)
λmax
then the µ -LMS algorithm is convergent in the mean.

117/120
µ -LMS Algorithm:
Convergence in the Mean (5)(2/2)
n

 Further, since tr [ R] = ∑ λ ≥ λ
i =0
i
(where tr [ R]
max

is the trace of R) convergence is assured if

2
0 <η < (76)
tr[ R ]

118/120
 Random Walk towards the Weiner
Solution
 Assume the
familiar fifth
19
95 141. 0.3
6
4 .56 4
12 141.88
39 8 839 16 41
21 41 66.1 117.6554 6.1
.3 1 12
0 682 11 4
4 19 93.42

order function
93 7 .
8 .42 655
982 68 4
69.1 69

4
2

55
2 5 .1
44.9697 98

7.6
39
28

µ -LMS uses a

11
.88

682
 0

141
Weiner solution

93.42682
93.42
w1 (0.339, -1.881)

828
-2

local estimate of

69.

69.19
µ LMS solution

44
5

982
(0.5373, -2.3311)
97

.9
-4

6
96

97
.

8
4

11

5
gradient to search
4

7.
-6 14

65
1 4

54
.8
8
93.
42 6 69.19828 55
39 82 7 .6
19 82
93 .426 11 8 3 9 2 4

the Weiner
-8 0
21 .341 16 1.8 11
4.5 6.1 117.6554 14 66.
12 1 1
69 4 4
-10 5 141.8839 9 0.3
-10 -5 0 5 1 10

solution in weight w0

space. 119/120
Adaptive Noise Cancellation

120/120