Bayesian inference is a probabilistic framework which permits decision-making in situations with incomplete information, maximally incorporating all available evidence, assumptions, and prior knowledge. Bayesian inference functions in a world of uncertainty and probabilities, which seems to mirror the issues of nanoelectronics. As neural networks, Bayesian models are highly topological in nature, suffer intensely of the von Neumann bottleneck, and map well to the concept of in-memory computing. This is why we strongly believe that novel nanoelectronic provide provide an ideal path for deploying Bayesian reasoning at the edge.
Bayesian models are not directly brain-inspired, but have been connected to biological intelligence, and significant evidence supports humans use of Bayesian inference for perception and other cognitive tasks. For example, visual illusions reveal the brain’s use of priors for image analysis, a signature of Bayesian inference. From the applicative side, Bayesian models have recently been used for applications such as robotics, multiple sensor data fusion, and computer-aided surgery. In fact, Bayesian models excel where neural networks struggle: due to their capability to incorporate prior knowledge and expertise, they deal well with small data situations, or when uncertainty is extreme. For these reasons, we are excited about their deployment at the edge.
We have taped out a first hybrid resistive memory / CMOS test chip implementing in-memory Bayesian inference, and are looking forward to testing it soon!