Spiking Neural Network Data Reduction via Interval Arithmetic

Approximate Computing (AxC) allows reducing the accuracy required by the user and the precision provided by the computing system to optimize the whole system in terms of performance, energy, and area reduction. Spiking Neural Networks (SNNs) are the new frontier for artificial intelligence because they better represent the timing influence on decision making, and also allow for a more reliable hardware design. Unfortunately, this design requires some area minimization strategies when the target hardware reaches the edge of computing. This seminal work introduces modeling of the approximation for data storage that supports an SNN via Interval Arithmetic (IA) by extracting the computation graph of the SNN and then resorting to IA to quickly evaluate the impact of approximation in terms of loss in accuracy without executing the network each time. Experimental results comparing our model to the real network confirm the quality of the approach