Bayesian inference using stochastic logic: A study of buffering schemes for mitigating autocorrelation

• Published in: International journal of approximate reasoning Vol. 112; pp. 4 - 21
• Main Author: Hoe, David H.K.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.09.2019
• Subjects: Autocorrelation mitigation; Bayesian inference; Muller C-element; Stochastic computation; Muller C-element; Bayesian inference; Stochastic computation; Autocorrelation mitigation
• ISSN: 0888-613X; 1873-4731
• DOI: 10.1016/j.ijar.2019.05.007

Bayesian inference has become near ubiquitous in the design of algorithms for machine learning and autonomous robotic systems as this method provides a rigorous mathematical framework for the probabilistically handling of data with elements of uncertainty. As conventional computer architectures are inefficient in their implementation of probabilistic approaches to inference, stochastic computation has recently emerged as a viable alternative. Recent research has demonstrated that Bayes' rule can be directly implemented with a Muller C-element using stochastic computing. However, the switching inertia at the output due to the inherent memory effect in the C-element limits the accuracy of this approach. Previous methods for reducing this autocorrelation effect have ranged in complexity from counter-based regeneration to bit-reshuffling with memory buffers. Simplified buffering techniques that function independently of the state of C-element are proposed in this work and the effectiveness of the autocorrelation mitigation is evaluated. In addition, the use of multi-input C-elements are shown to add design flexibility for implementing Bayesian inference. Detailed numerical simulations and analysis of configurations that range from a simple cascade of gates to tree structures with large multi-input C-elements demonstrate the viability and limitations of the proposed simplified buffering approaches. These results are significant for their potential for enabling compact implementations of stochastic Bayesian approaches as control lines and random number generators can be shared across multiple C-elements.