Fully-Integrated Spiking Neural Network Using SiOx-Based RRAM as Synaptic Device

This paper presents, to the best of the authors’ knowledge, the first complete integration of a Spiking Neural Network combining analog neurons and SiO x -based resistive memory (RRAM) synapses. The implemented topology is a perceptron, and the circuit is aimed at performing MNIST digits classification. An existing framework was adapted for off-line learning and weight quantization, and the network was later converted into its spiking equivalent. The test chip, fabricated in 130 nm CMOS, shows a classification accuracy of 82%, with a 180 pJ energy dissipation per spike.