Memristor-Based Binary Synapses for Deep Neural Networks

Abstract:
The development of biologically-oriented mathematical models has allowed recent advances in neuromorphic computing architectures and in the understanding of the mechanisms behind the complex dynamics of living systems. Deep Neural Networks are among the most computational efficient architectures used in machine learning. The simplest structure is represented by multiple-layers perceptrons with binary synapses (i.e. the synaptic weights assume binary values). The manuscript introduces a memristor-based circuit to implement an artificial binary synapse. In the paper it will be shown how the binary output is obtained with respect to the internal state of the memristor and how this kind of sub-system could be a more efficient implementation of synapses inside networks such as a perceptron.