Learning method for ex-situ training of memristor crossbar based multi-layer neural network

Abstract

Memristor is being considered as a game changer for the realization of neuromorphic hardware systems due to its similarity with biological synapse. Recent studies show that memristor crossbar can provide high density and high performance neural network hardware implementation at low power due to its physical layout, nano scale size and low power consumption feature. This paper describes the training method that can be used for the implementation of memristive multi-layer neural network with ex-situ method. We mimic the behavior of memristor crossbar in software training process to achieve more accurate and close computations to hardware. Voltage divider has been used to calculate the dot product in this method. To demonstrate the accuracy and effectiveness of this method, different patterns and non-separable functions using memristor crossbar structures are simulated. The results demonstrate that more accurate computations can be produced using this learning method for ex-situ. It also reduces the learning time of functions.

A Highly Robust Binary Neural Network Inference Accelerator Based on Binary Memristors

“…The traditional differential pair structure is shown in Figure 6a, where the conductance difference of two memristors (a differential pair) is used to represent one synapse, each of the two memristors has two resistive states: high-resistance state (HRS) and low-resistance state (LRS), the corresponding conductance states are denoted as G HRS and G LRS . This is a classic scheme to map positive and negative weights in memristive crossbar array, which is also used in previous work [39][40][41][42]. Differential operation is performed on the two memristors to achieve mapping of the two weight values of +1 and −1, where the output feature, namely the current Ik of each column of memristor array, is expressed as:…”