On the maximum function in stochastic computing

Abstract

Stochastic circuits (SCs) offer significant area, power and energy benefits at the cost of computational inaccuracies. SCs have received particular attention recently in the context of neural networks (NNs). Many NNs use the maximum function, e.g., in the max-pooling layer of convolutional NNs. Currently, approximate workarounds are often employed for this function. We propose NMax, a new SC design for the maximum function that produces an exact result with latency similar to an approximate circuit. Furthermore, unlike most stochastic functions, NMax is correlation insensitive. We also observe that maximum calculations are subject to application-specific bias and analyze this bias.