Fault Recovery in Spiking Neural Networks Through Target and Selection of Faulty Neurons for 3D Spiking Neuromorphic Processors

Neuromorphic systems integrate features inspired by the neurobiological system which makes them suitable for many cognitive applications like facial recognition, autonomous driving, and bio-medicine. SNN is implemented in these systems for asynchronous information processing with energy efficiency. Many cognitive applications, however, require large numbers of highly interconnected neurons which are often noisy as a result of interference from both internal and external sources. Due to this, these neurons develop faults over time and their computation output becomes unreliable. In this paper, first, we investigated the effect of faulty neurons in spiking neural networks (SNNs). Afterward, we propose a recovery method ’target and selection’ (TSM) for fault recovery in terms of model classification accuracy. In SNN models with stuck-at (SA) faults, TSM identifies and removes all faulty neurons less complexly. Evaluation results show that with the proposed fault recovery method, we could achieve a classification accuracy loss of less than 4% in SA faulty SNNs far less than using the proposed previous fault recovery method at a high fault rate.