You are on page 1of 7

Biological Neural Networks 046041: Simulation Project

July 4, 2017

Jacob Minsky
yaakov34@campus.technion.ac.il - 317912228
Jacob Minsky Biological Neural Networks 046041: Simulation Project

1 Code Description

The simulator uses a Leaky Integrate and Fire model to simulate the ganglion cells and the V1 horizontal and vertical
detection cells.

[Lines 65-66] At each time step, the input image is convolved with a 3x3 matrix to produce the values seen by the
receptive fields of the ganglion cells; the matrix has a high positive central value and smaller negative peripheral values
with a total sum of 0, so that there will be no response to spatially uniform input. The convolved image is presented to
center ON ganglion cells, and its negative to center OFF cells; the two arrays of ganglion cells correspond in size to
the input field without its outermost edge (the receptive fields of ganglions at the end would be incomplete).

[Lines 57-61] The ganglion cells whose membrane voltage has exceeded a threshold are located and a spike is recorded
for them; the voltages for these cells are reset to vr .

[Lines 76-82] In the same manner, the spikes are detected for the v1h and v1v cells.

[Lines 84-90] The voltages are computed for the v1v and v1h visual cortex cells, based on the spikes sent by the
ganglion cells. There is a vector of ganglion v1h cells and a vector of v1v cells (the drawing in the handout is slightly
different, showing a full array of v1v and v1h cells). Horizontal detection cells integrate spikes from rows of ganglion
cells, and vertical detection cells integrate spikes from columns of ganglion cells; the sum is taken in such a way that
each signal from an ON ganglion column is summed with the signal from the neighbouring OFF columns. In other
words, the receptive field of each v1v cell consists of three inputs, each of which is the sum of a column of ganglion
cells: the central input is the sum of the column of ON ganglion cells, and the peripheral inputs are the sums of
columns of OFF cells to each side of the ON column.

Lateral suppression works to reduce the activity of v1v cells when v1h cells are active, and vice versa.

2 Features not modeled

The model includes only one type of retinal neurons: the ganglion cells. It does not include

bipolar and amacrine cells which transfer information between the rods and cones and the ganglion cells (the
innervation pattern is what forms the ON and OFF receptive field).

horizontal cells, which provide inhibitory feedback to rods and cones, and enable background suppression and
adjustment to different light levels.

modeling of the fovea. The neural organization is different in the fovea and includes a 1:1 map of the received
image.

The modeled visual cortex includes only orientation detection cells; these are part of the the early vision system, while
the higher layers of the visual cortex process more complex features. The temporal frequency response is not modeled;
the visual cortex tends to respond more to changing stimuli.

Page 1 of 6
Jacob Minsky Biological Neural Networks 046041: Simulation Project

3 Behavior of a ganglion cell

The following is the spike-triggered average image for a specific ON ganglion cell at delays of 4,8,12,16, and 20 ms.
The structure of a high center and low values around it is clearly visible; the rest of the imput is simply averaged noise,
since spikes of only one neuron were considered. This confirms the shape of the receptive field of the neuron. A very
similar picture with reversed colors is obtained for an OFF ganglion.

2 2 2 2 2

4 4 4 4 4

6 6 6 6 6

8 8 8 8 8

2 4 6 8 2 4 6 8 2 4 6 8 2 4 6 8 2 4 6 8

Figure 1: Spike-Triggered Average

The rise in the membrane potential leading to a spike, produced directly from the LIF equation, is as follows:

40

35

30
Membrane Potential[mV]

25

20

15

10

0
0 50 100 150
Time [s]

Figure 2: Membrane potential

The reset to vr after a spike is visible, as is the charging behavior of the RC circuit.

4 Spiking behaviour of a neuron

The following is a spike train of a glial neuron from an experiment:

Page 2 of 6
Jacob Minsky Biological Neural Networks 046041: Simulation Project

spike train
2

1.5

0.5

0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000

Figure 3: Spike train of an RGC

The Gaussian smoothing is plotted under the spike train in red. I found an optimal kernel for the first trial of this
experiment, and applied it to the rest of the data. The optimal kernel found was as follows:

0.025

0.02

0.015

0.01

0.005

0
0 20 40 60 80 100

Figure 4: Optimal Kernel

It is very close to a Gaussian filter, and results of a convolution with it are hard to distinguish from the original Gaussian
smoothing in a plot.

Page 3 of 6
Jacob Minsky Biological Neural Networks 046041: Simulation Project

5 Horizontal and vertical orientation detection by neurons

2 2 2 2
4 4 4 4
6 6 6 6
8 8 8 8

2 4 6 8 2 4 6 8 2 4 6 8 2 4 6 8

2 2 2 2
4 4 4 4
6 6 6 6
8 8 8 8

2 4 6 8 2 4 6 8 2 4 6 8 2 4 6 8

2 2 2 2
4 4 4 4
6 6 6 6
8 8 8 8

2 4 6 8 2 4 6 8 2 4 6 8 2 4 6 8

Figure 5: Stimuli presented to the simulated retina

These stimuli are anti-aliased lines of width 1 in a 9x9 input matrix (I found that anti-aliasing the lines results in much
less noisy output). The stimuli were presented to the retina in 10 trials, resulting in neuron behavior such as this:

Page 4 of 6
Jacob Minsky Biological Neural Networks 046041: Simulation Project

Figure 6: Neuron activity

The periods when the stimulus was on/off are clearly seen. The tuning curve for the V1 neurons was as follows:

1400

1200

1000

800

600

400

200

0
0 2 4 6 8 10 12

Figure 7: Tuning curves of vertical (red) and horizontal (blue) V1 neurons with lateral inhibition

Page 5 of 6
Jacob Minsky Biological Neural Networks 046041: Simulation Project

1600

1400

1200

1000

800

600

400

200
0 2 4 6 8 10 12

Figure 8: Tuning curves of vertical (red) and horizontal (blue) V1 neurons without lateral inhibition

The curve with inhibition has a steeper falloff with angle (roughly hyperbolic shape instead of roughly linear shape of
the rise), and would be better suited for distinguishing values closer to horizontal and vertical orientations. The no-
inhibition curves would be suited to full-range orientation determination. The model is able to distinguish orientations
only within a 90-degree sector: the curves are symmetric about the vertical orientation.

Noise added to the model results in a lower contrast ratio of the curves.

6 Decoding of V1 neuron output

I used a population-count algorithm: from the firing V1 neurons, the most typical (median) neurons are chosen, and
their average rate determined. The ratio of the rates is then taken and looked up in a table of ratios of horizontal and
vertical firing rates. For this model, this results in very good decoding performance.

Page 6 of 6

You might also like