B²N²: Resource efficient Bayesian neural network accelerator using Bernoulli sampler on FPGA

Abstract

A resource efficient hardware accelerator for Bayesian neural network (BNN) named B²N², Bernoulli random number based Bayesian neural network accelerator, is proposed. As neural networks expand their application into risk sensitive domains where mispredictions may cause serious social and economic losses, evaluating the NN’s confidence on its prediction has emerged as a critical concern. Among many uncertainty evaluation methods, BNN provides a theoretically grounded way to evaluate the uncertainty of NN’s output by treating network parameters as random variables. By exploiting the central limit theorem, we propose to replace costly Gaussian random number generators (RNG) with Bernoulli RNG which can be efficiently implemented on hardware since the possible outcome from Bernoulli distribution is binary. We demonstrate that B²N² implemented on Xilinx ZCU104 FPGA board consumes only 465 DSPs and 81661 LUTs which corresponds to 50.9% and 14.3% reductions compared to Gaussian-BNN (Hirayama et al., 2020) implemented on the same FPGA board for fair comparison. We further compare B²N² with VIBNN (Cai et al., 2018), which shows that B²N² successfully reduced DSPs and LUTs usages by 50.9% and 57.9%, respectively. Owing to the reduced hardware resources, B²N² improved energy efficiency by 7.50% and 57.5% compared to Gaussian-BNN (Hirayama et al., 2020) and VIBNN (Cai et al., 2018), respectively.