Stochastic Mean Circuits Based on Inner-Product Units Using Correlated Bitstreams

Yongqiang Zhang, Xiaoyue Chen, Jie Han, Guangjun Xie

Abstract: Stochastic mean circuits (SMCs) are at the core of mean filtering and neural networks, but they have not been fully investigated in stochastic computing (SC). In this brief, SMCs in the unipolar and bipolar formats are respectively proposed to average bitstreams. For this purpose, the unipolar and bipolar stochastic inner-product circuits (SIPCs) are first designed by using positively correlated bitstreams for multiplication and accumulation. These bitstreams are generated by sharing a random number source (RNS) among different stochastic number generators (SNGs) to lower hardware costs. Applications of the proposed designs in Gaussian filtering, edge detection, and mean filtering are implemented. Experimental results show that the proposed SIPCs and SMCs outperform previous designs in both hardware efficiency and computing accuracy.