The robustness of Bayesian neural networks (BNNs) to real-world uncertainties and incompleteness has led to their application in some safety-critical fields. However, evaluating uncertainty during BNN inference requires repeated sampling and feed-forward computing, making them challenging to deploy in low-power or embedded devices. This article proposes the use of stochastic computing (SC) to optimize the hardware performance of BNN inference in terms of energy consumption and hardware utilization. The proposed approach adopts bitstream to represent Gaussian random number and applies it in the inference phase. This allows for the omission of complex transformation computations in the central limit theorem-based Gaussian random number generating (CLT-based GRNG) method and the simplification of multipliers as and operations. Furthermore, an asynchronous parallel pipeline calculation technique is proposed in computing block to enhance operation speed. Compared with conventional binary radix-based BNN, SC-based BNN (StocBNN) realized by FPGA with 128-bit bitstream consumes much less energy consumption and hardware resources with less than 0.1% accuracy decrease when dealing with MNIST/Fashion-MNIST datasets.

Bayesian neural network (BNN) has gradually attracted researchers' attention with its uncertainty representation and high robustness. However, high computational complexity, large number of sampling operations, and the von-Neumann architecture make a great limitation for the further deployment of BNN on edge devices. In this article, a new computing-in-MRAM BNN architecture (CiM-BNN) is proposed for stochastic computing (SC)-based BNN to alleviate these problems. In SC domain, neural network parameters are represented in bitstream format. In order to leverage the characteristics of bitstreams, CiM-BNN redesigns the computing-in-memory architecture without complex peripheral circuit requirements and MRAM state flipping. Additionally, real-time Gaussian random number generators are designed using MRAM's stochastic property to further improve energy efficiency. Cadence Virtuoso is used to evaluate the proposed architecture. Simulation results show that energy consumption is reduced more than 93.6% with slight accuracy decrease compared to FPGA implementation with von-Neumann architecture in SC domain.