Implementation of evolutionary learning hardware realized in a TaOx-based memristive binary neural network via stateful logic

Abstract

Memristors are attracting as synaptic devices for neuromorphic computing because of their resistive switching and non-volatility. However, their non-idealities such as non-linearity, asymmetry, and variability degrade the performance of a memristive neural network. In order to deal with the device non-idealities, a memristive binary neural network that uses only two resistance states of the memristors, can be used. Hardware-level training of the memristive binary neural network is not efficient since it requires a peripheral circuitry for the operation of the conventional optimization algorithm and an external memory for storing the intermediate operation results. In this study, we propose a technology to implement a genetic algorithm, an optimization algorithm inspired by the principle of evolution, on a memristor crossbar array. For the technology, the stochastic nature of memristor materials and stateful logic technology was utilized. In addition, several evolutionary learning of the memristive binary neural network using the proposed genetic algorithm is performed by simulation. The simulated classification accuracy is 90.1 % which is equivalent to the accuracy (90 %) of the conventional training method, suggesting validity of the proposed technology.