Biomedical Engineering Reference
In-Depth Information
limited only by the overall bandwidth (i.e., the multiplexing rate) of the system.
Connections to synapses can be programmed to have one of 30 different
weights, 15 inhibitory and 15 excitatory (24). During activation of a synapse, its
conductance changes from an essentially nonconducting state to one of the 30
conductance values for a duration of 50 ns. Weights for each synapse are stored
off-chip in a connection list along with the synapse address, thereby allowing
multiple connections with independent weights to the same compartment.
The present system shown in Figure 2 holds 128 neuromorphs. However,
with higher-density neuromorph chips it is capable of simultaneously intercon-
necting and running over 1000 neuromorphs. The entire network connectivity
can be changed in a fraction of a second, allowing rapid evaluation of a large
number of different architectures. For this purpose, the spiking activity from a
selected set of neuromorphs can be sampled at any time.
4.
NEUROMORPHS IN A WINNERLESS COMPETITION NETWORK
The coordinated firing of neurons in a network is likely to be a key feature
of neural computation. Interconnected neurons with steady input behave as cou-
pled oscillators, and such systems can display a range of coordinated behaviors.
The phenomenon of synchronous firing of groups of neurons (8) is probably the
best known example of this. Other, more complex patterns of spikes can result
from the interaction between neurons—the behavior depends on the details of
the coupling. Such networks are strong candidates for computational mecha-
nisms that biological systems may actually use for representation and recogni-
tion of patterns (16). One that has attractive features for application to small
networks of neuromorphs is the winnerless competition network of Rabinovich
et al. (23). In the absence of a stimulus, these networks are quiescent or display
unpatterned activity, but when stimulated they generate cyclic patterns of spik-
ing activity that are distinct and characteristic of the stimulus. These networks
achieve spatiotemporal coding by executing heteroclinic orbits around saddles in
a space that changes when the pattern of stimuli changes. The behavior, which
captures some features observed of olfactory processing in the locust antennal
lobe, was generated in a simulation with nine model neurons (FitzHugh-
Nagumo) with strong asymmetric inhibitory interconnections.
We have produced similar behavior in a network of neuromorphs intercon-
nected and stimulated similarly to the model neurons in the simulation of Rabi-
novich et al. The excitability of each neuromorph in the network is controlled by
(a) supplying a 100-Hz spike train to the "lower" threshold synapse (Figure 1A)
and (b) feeding back the neuromorph's output spikes to the "upper" threshold
synapse. The ratio of "upper" to "lower" spikes determines the spike-firing
threshold; the negative feedback limits the firing rate, giving the desired back-
ground behavior (see (6)). The temporal pattern of spiking of the neuromorphic
