Conference Proceedings of the IEEE Magnetics Society Skyrmionic Synapse Implementation for Pattern Recognition Using Convolutional Neural Network

Conventional neuromorphic computing faces the foreseeable limit of computational and power resources that can be mitigated by using spintronic-based devices. In this digest, we propose a 4-bit (16-state) skyrmion-based synaptic device for application in Convolutional Neural Network(CNN) consuming 0.971 fJ energy per weight update. Our proposed device offers compactness and scalability to higher bits. The skyrmionic synapse device is simulated using a micromagnetic simulation package(OOMMF) integrated with a software-based ReLU Max Pooled function to implement CNN for neuromorphic computing applications. We attained an accuracy of 98.15 % for pattern recognition on the MNIST handwritten data set.