VCAM: Variation Compensation through Activation Matching for Analog Binarized Neural Networks

We propose an energy-efficient analog implementation of binarized neural network with a novel technique called VCAM, variation compensation through activation matching. The architecture consists of 1T1R ReRAM arrays and differential amplifiers for implementing synapses and neurons, respectively. To restore classification test accuracy degraded by process variation, we adjust the biases of the neurons to match their average output activations with those of ideal neurons. Experimental results show that the proposed approach recovers the accuracy to 98.55% on MNIST and 89.63% on CIFAR-10 even in the presence of 50% threshold voltage and 15% resistance variations at 3-sigma point. This result corresponds to the accuracy degradation of only 0.05% and 1.35%, respectively, compared to the ideal case.