Neuromorphic Synapses for Artificial Dendrites

Wayne C. Westerman, David P. M. Northmore, and John G. Elias

University of Delaware

Newark, DE. 19716

ABSTRACT

Neuromorphic, variable-weight synapses on artificial dendrites are described that facilitate experimentation with correlative adaptation rules. Attention is focused on those aspects of biological synaptic function that likely have an impact on a neuromorphic network's computational power and adaptive capability. These include sublinear summation, quantal synaptic noise, and independent adaptation of adjacent synapses. For present purposes, storing synaptic weights off chip simplifies the addition of quantal weight noise and allows connections to the same dendritic compartment from different sources to have independent weights. Neuromorphic synapses are implemented as conductances to mimic biological synapses and thus enable sublinear summation. The diffusive nature of artificial dendrites is shown to add flexibility to the design of fast synapses by allowing the duration that conductances are enabled to be very short or variable. We present two complementary synapse designs, the shared conductance array and the self-timed synapse. The former achieves weight variation by selecting different conductances from an on-chip array, and the latter by modulating the length of time a constant conductance remains activated. Both work with our interchip communication system, virtual wires.