Verification of neuromorphic processing accuracy against non-ideal errors in synapse-based neural network systems

Abstract

The synapse-based neuromorphic systems for deep neural network (DNN) require neuron read/update circuit blocks which specifications depend on the synapse components. In this study, we implemented and verified the dedicated circuit system to operate the three-terminal analog synaptic memory cells, electrochemical random access memory (ECRAM), for multi-bit analog neuromorphic computing. In addition, we analyzed the impact of the noise/offset generated by the neuron circuit systems and synapse cell arrays, which can affect the neuromorphic processing accuracy, by using the Modified National Institute of Standards and Technology database (MNIST) dataset. The experiments with MNIST datasets were conducted in two ways: by performing ideal inference simulations with MATLAB and experiments with the neuromorphic system board. The results of the ideal inference simulation and experiment were 97.4% and 96.92%, respectively. The accuracy of 97.31% was measured when the weight of the hidden layer was set with the fixed resistance values to confirm the effectiveness of the synapse cells. According to the results, the effects of synapse cells and neuron circuits to the processing accuracy were 0.09% and 0.39%, respectively. Also, this MNIST experiments verified that about 10% or smaller variations of weight values in the synapse cells lead to negligible effects on processing accuracy through the training and inference. The figure above illustrates the read/update circuit blocks for an ECRAM synapse-based neuromorphic system designed for deep neural networks (DNNs). Analyzing the impact of noise/offsets originating from the above circuit system using the MNIST dataset, influence of the synapse cell and neuron circuit on accuracy was 0.09% and 0.39%, respectively. Through training and inference, variations of less than 10% in the synaptic cell's weight values have a minimal impact on processing accuracy. image