Variation-Tolerant Capacitive Array for Binarized Neural Network

In this letter, we propose the variation-tolerant capacitive binarized neural network using the capacitance of 2-terminal MANOS memory devices. The capacitance-voltage characteristic of the synaptic device is shifted by program or erase operation. Two saturated regions of the C-V curve are used for representing the weight of the binarized neural network (BNN), 1 and -1. Thus, the capacitance variation of the proposed synaptic device is much less than the conductance variation of the conventional synaptic devices. The effects of V-th variation and memory window on accuracy are analyzed with CIFAR10 datasets. Our simulation result shows that the proposed capacitance-based BNN is much more robust to V-th variation compared to the conductance-based BNN.

Automated summary

This letter proposes a variation-tolerant capacitive binarized neural network using 2-terminal MANOS memory devices. The capacitance-voltage characteristic of the synaptic device is used to represent the weight of the binarized neural network. The proposed synaptic device shows less capacitance variation compared to conventional synaptic devices, resulting in a more robust network.