A SiOx, RRAM-based hardware with spike frequency adaptation for power-saving continual learning in convolutional neural networks

Abstract

Biological systems autonomously evolve to maximize their efficiency in a continually changing world. On the other hand, artificial neural networks (ANNs) outperform the human ability of object recognition but cannot acquire new information without forgetting trained tasks. To introduce resilience in ANNs, we present a SiOx RRAM-based inference hardware capable of merging the efficiency of convolutional ANNs and the plasticity of spiking networks. We validate the accuracy of the system with MNIST (99.3%), noisyN-MNIST (96%), Fashion-MNIST (93%) and CIFAR-10 (91 %) datasets. We demonstrate that the circuit plastically adapts its operative frequency for power saving and enables continual learning of up to 50% non-trained classes. This optimizes the classification and enables the re-training of the filters, thus overcoming the catastrophic forgetting of standard ANN s.