Fast Exploration of the Impact of Precision Reduction on Spiking Neural Networks

Approximate Computing (AxC) techniques trade off computational accuracy for gains in performance, energy efficiency, and area reduction. This trade-off is particularly advantageous when applications, like Spiking Neural Networks (SNNs), are naturally tolerant to some degree of accuracy loss. SNNs are especially practical when the target hardware is pushed to the edge of its computing capabilities, necessitating area minimization strategies. In this work, we utilize an Interval Arithmetic (IA)-based model that propagates approximation errors through the application’s computation flow to assess these approximations' impact on the outputs. We enhance this IA-based model by introducing observation points within the computation flow to quickly detect when the level of approximation surpasses a set threshold. Experimental results demonstrate the model’s effectiveness in significantly reducing exploration time, enabling more precise and fine-grained approximations that further minimize network parameters.