Skyrmionic implementation of Spike Time Dependent Plasticity (STDP) enabled Spiking Neural Network (SNN) under supervised learning scheme

Abstract

Hardware implementation of Artificial Neural Network (ANN) algorithms, which are being currently used widely by the data sciences community, provides advantages of memory-computing intertwining, high speed and low energy dissipation which software implementation of the same does not have. In this paper, we simulate a spintronic hardware implementation of a third generation neural network - Spike Time Dependent Plasticity (STDP) learning enabled Spiking Neural Network (SNN), which is closer to functioning of the brain than most other ANN-s. Spin orbit torque driven skyrmionic device, driven by a transistor based circuit to enable STDP, is used as a synapse here. We use a combination of micromagnetic simulations, transistor circuit simulations and implementation of SNN algorithm in a numerical package to simulate our skyrmionic SNN. We train the skyrmionic SNN on different datasets under a supervised learning scheme and calculate the energy dissipated in updating the weights of the synapses in order to train the network.