A NEUROMORPHIC SYSTEM - Complex Systems Science in Biomedicine - page 789

Figure 6 . Weight matrices connecting ( A ) whiskers and ( B ) visual sustained units to the left
and right Hebb Layers. Excitatory weights are shown in white, inhibitory in black. Height of
bars represent weight values. Inputs in the whiskers and visual space (-60 to +60() are repre-
sented along rows; the different Hebb layer units along columns.
form an inhibitory surround. If the spiking activity of the Hebb layers is now
connected to the vehicle's steering, the vehicle avoids obstacles visually, without
contact with the whiskers.
Negotiating a complex environment requires that information other than just
luminance is brought to bear; judging distances, reacting to impending collisions
and guiding locomotion require the use of visual motion cues. The network of
Figure 4 and the application of Hebbian mechanisms during sensorimotor ex-
perience results in the development of motion selective units. For this, we added
transient visual units that fire to changes in luminance. Because the extraction of
velocity information depends on correlations between stimulus positions at dif-
ferent times, the spikes from the transient units are connected to the Hebb layers
via different time delays. For the sake of illustration, the transient unit array is
connected to the Hebb layers via four weight matrices, one each for delays of 0,
0.5, 1.0, and 1.5 ms (Figure 7A,B). In a neuromorphic network, these delays
could be implemented by the facility for programmable axonal delays in the
Virtual Wires system, or by dendritic delays that depend on the site of synaptic
activation along a dendrite. Under appropriate conditions, the application of the
Hebb rule in the presence of a moving stimulus image could selectively
strengthen inputs that come from the four positions in space occupied by the
image at times t , t - 0.5, t - 1.0, and t - 1.5 ms. For Hebb layer units to come to
respond to the motion, they must be firing at time t . Therefore, they require an
"instructive" input generated during image movement to bias them on so that
they fire. In this example, the image motion is self-produced, as by eye or head