Neuromorphic Synapses for Artificial Dendrites

Abstract

In the past few years several researchers have developed general-purpose spiking silicon neurons, or neuromorphs, which attempt to capture various functional behaviors of biological neurons [19, 23, 24]. These CMOS analog VLSI implementations are intended to be the computational elements linking sensory input to motor output in artificial nervous systems. Systems consisting of thousands or millions of neuromorphs may be required to realize certain basic behaviors. However, until appropriate methods are found to specify and adapt network parameters (e.g., connectivity, synapse weights, dynamics, morphology), these systems will remain beyond our grasp. Though various Hebbian or correlative adaptation rules have been applied to perceptron based networks (e.g., [77, 11, 141]) and single-compartment model neurons [12, 26], these rules are virtually untested with multi-compartment spiking neuromorphs that model the spatially extensive dendrites found in biological systems. This paper discusses the VLSI design and performance of just one of the adaptive network parameters: variable synaptic conductances that mimic the fast chemical synapses found on dendrites throughout high-order nervous systems.