A Reliability Assessment Approach for A LIF Neurons Based Spiking Neural Network Circuit

Although Spiking Neural Networks(SNN) show high robustness, they are inevitably affected by external factors or within the devices themselves when transferring them into hardware circuits. For example, devices in hardware circuits may cause physical defects, aging, or catastrophic failures during the manufacturing process, which can degrade the performance of SNN circuits. To investigate these effects, this study selects existing circuits and simulates the effects on the performance of SNN circuits under different failure scenarios. By analysis the recognition accuracies as functions of the level of failure and degradation, the impact of circuit failures on the performance of the selected SNN can be evaluated. It is shown that the second hidden layer is very sensitive to noise and neuron failures, and therefore more reliable circuits or devices are needed to improve circuit reliability. In the second hidden layer, the loss of about 85.25% of the neuron circuits does not significantly affect the performance of the whole SNN, which indicates the degeneracy of SNN hardware circuits. The connections between neurons can be reduced by layer-by-layer, global and local pruning to reduce the cost and optimize the circuit.SNN hardware circuits can tolerate up to 93.33% fault tolerance, which is an important guideline for the application of SNN hardware circuits.