SCFNNs: Stochastic Computing Fuzzy Neural Networks

The resource intensity of fuzzy neural networks (FNNs) may hinder their ubiquity, particularly in resource-constrained hardware platforms such as edge devices in Internet of Things (IoT) applications. In this letter, stochastic computing fuzzy neural networks (SCFNNs) are proposed as hardware-efficient FNNs to mitigate the resource demand. In SCFNNs, stochastic computing (SC) is tailored to implement FNNs, rather than using conventional floating-point (FP) and fixed-point (FxP) representations. By leveraging the compatibility of SC and fuzzy arithmetic, range adjustment as required by conventional SC can be relaxed. Experiments on Iris, Wine, and Breast Cancer datasets, comparing FP, FxP, and SC implementations, manifest the advantage of the SC design in terms of hardware efficiency and performance in most cases, without degradation in accuracy.