Biomedical Engineering Reference
In-Depth Information
Figure 2
. Domain board. Dimensions: 14.5 cm
2
. The top half of the board is occupied by
8 chips (0.5-
m CMOS (complementary metal oxide semiconductor)), each containing
16 neuromorphs, each of which containing 4 dendritic branches. Most of the 2.2 + 2.2
mm chip area is filled by dendrites. The lower half of the board contains the spike-
routing system, connection memories, and connectors for a host computer.
the activated or on state, and (b) by the duration of the impulse supplied to the
transistor's gate. Activation of excitatory and inhibitory synapses at different
sites set up potentials that diffuse over the dendritic tree. The potential at the
point where the branches converge (Figure 1A) affects the rate of output spike
firing. The dynamics, which determine the dendritic delays, are set by the pro-
grammable membrane and axial resistances (R
m
) and (R
a
), and by the fixed-value
compartmental capacitance (5). Whether the neuromorph fires or not and its rate
of firing depend upon whether and how much the branch-point potential exceeds
a firing threshold implemented by a comparator in the "soma" (Figure 1A). For
this threshold potential to be influenced by the activity of other neuromorphs in
a network, it is set by the ratio of spike frequency applied to two special syn-
apses, the "upper" and "lower" in Figure 1A (6).
To build a network, one must provide a means for sending the output spikes
of any neuromorph to the synapses of any other. In addition, the spikes must
arrive at their destinations in a time that is short compared to network dynamics.
To this end, we developed a multiplexing scheme that we call "Virtual Wires"
(4). As with Mahowald's (12) method of connecting neuron outputs to synapses,
addresses are multiplexed in time over a few dedicated wires rather than through
a much larger number of direct connections. In our scheme, a spike generated by
a neuromorph activates destination synapses with delays that are programmable
for each connection. The number of connections each neuromorph may have is
