Biomedical Engineering Reference
In-Depth Information
Figure 7 shows examples of how weights from the transient units to the
Hebb layer units developed. The four weight matrices at increasing delays show
progressive shifts in their receptive fields, thereby generating a skewed spatio-
temporal receptive field required for directional selectivity (1). To test direc-
tional selectivity, a single visible object was rotated at a constant rate through
the visual field. Figure 7C,D (lower half) shows that Hebb Layer Left responded
with spikes to leftward motion but not to rightward motion, and vice versa for
Hebb Layer Right. Note also that the units have restricted receptive fields organ-
ized topographically, a first step towards analyzing motion flow fields. We can
now envision some further steps that would be required to use this information
for behavioral guidance: the development of selectivities for different velocities
and different patterns of motion as exhibited by neurons in the visual areas of
the cerebral cortex.
7.
NEUROMORPHS IN NEURAL PROSTHETICS
Neuromorphic systems of the type we are developing may have a special
advantage for neural prosthetics in that neuromorphs naturally deal with spike
signals: they accept spike inputs and they generate spike outputs. A very small
neuromorphic network implanted in a paralyzed patient could interpret spike
signals from an array of electrodes embedded in the motor cortex (20). The net-
work would extract the relevant information contained in the frequency and tim-
ing of the cortical spikes and would generate output spike trains patterned so as
to activate neurons, fiber tracts, or muscles, restoring lost function to the patient.
Neuromorphic prosthetics will certainly require the ability to adapt and learn
because the input spikes are largely arbitrary in nature—they are whatever the
implanted electrodes can pick up from the CNS—and they must be associated
with patterns of spike outputs for controlling behavior (22). While our neuro-
morphic development system (domain board) can implement arbitrary learning
schemes, it currently requires the assistance of an external computer.
8.
CONCLUSIONS
The challenge to building neuromorphic systems is deciding what features
to incorporate into the neuromorphic units and how to connect them to perform
usefully. A process of design from biophysical principles is an estimable ap-
proach but difficult (7). Faced with the great variety of potential mechanisms
that neurobiological research is revealing and the complex, nonlinear interac-
tions between them, we prefer to seek neuromorphic systems that are relatively
simple to make and able to self-organize and adapt dynamically. The results
presented here show that valuable capabilities can emerge in networks through
