Hardware implementation of Bayesian decision-making with memristors

Abstract

Brains perform decision-making by Bayes theorem – events are quantified as probabilities and based on probability rules, computed to render the decisions. Learning from this, Bayes theorem may be applied to enable efficient user–scene interactions. However, given the probabilistic nature, implementing Bayes theorem with the conventional deterministic computing hardware can incur excessive computational cost and decision latency. Though challenging, here a probabilistic computing approach is presented based on memristors to implement Bayes theorem. Memristors are integrated with Boolean logic circuits and, by exploiting the volatile stochastic switching of the memristors, realize probabilistic Boolean logic operations, key for Bayes theorem hardware implementation. To empirically validate the efficacy of the hardware Bayes theorem in enabling user–scene interactions, lightweight Bayesian inference and fusion operators are designed using the probabilistic logic circuits and apply the operators in road scene parsing for self-driving, including route planning and obstacle detection. The results show the operators can achieve reliable decisions in less than 0.4 ms (or equivalently 2500 fps), outperforming human decision-making and the existing driving assistance systems.