Implementation of multi-layer neural network system for neuromorphic hardware architecture

Abstract

We propose a new neuromorphic hardware system that is optimized to implement a multi-layer guide training algorithm, which is a kind of reinforcement training algorithm. To consider the hardware implementation, we apply the guide training algorithm that is simple and very suitable for memristor synapse. The system is modeled using Simulink and the accuracy of the system is verified by classifying 'T', 'X', and 'V' in 3x3 letter image. The target image of hidden layer is set to the inverted image of the input image. Using this proposed system architecture, the reinforcement learning in multi-layer can be easily implemented in hardware. © 2019 Institute of Electronics and Information Engineers (IEIE).