Machine Learning-Assisted Analysis of Advanced STDP for Neuromorphic Computing using MRAM

A spiking neural network (SNN) comprises spiking neurons that mimic the information transfer in biological neurons using a series of time-dependent spikes. The spikes from such neurons are sparse in time and space, and event-driven. This facilitates the development of low-power neuromorphic hardware when coupled with bioplausible local spike-timing-dependent plasticity (STDP) learning algorithm that can encode temporal information to solve complicated time-dependent pattern recognition problems. The SNN hardware implementation using novel memristive (spintronic) devices is an active area of research to achieve area and power efficiency. Spintronic devices pave for hardware-efficient implementation of complex neuromorphic algorithms such as STDP for in-situ learning. This work presents the implementation of STDP algorithms using spin-based synaptic devices. Moreover, using an unsupervised learning scheme, the paper presents an SNN for digit recognition that is based on the mechanism with increased biological plausibility with 3 different STDP learning rules. The results show the testing accuracy of the triplet-based STDP rule is 18.34%, 10.72%, 7.15%, and 3.53% higher than the pair-based STDP rule with 100, 250, 400, and 1600 excitatory neurons respectively. The energy with the presented device is reported as 300 fJ.