Supporting AI models that quantify uncertainty with emerging logic and memory technologies.

As the number of sensors and data collection streams employed by the DoD and its partners increases, machine learning (ML) has become increasingly invaluable for its ability to extract useful information from a massive influx of data. In the pursuit of faster decision times, removing humans from the decision loop is inevitable; however, ML algorithms are unpredictable and prone to undesired effects. Explainable AI--ML algorithms that can quantify the certainty of their decisions--offer a solution by allowing the user to set a confidence threshold where human intervention would be necessary. Bayesian neural networks (BNNs) are one such algorithm of interest, but they impose a significant computational overhead associated with random number generation (RNG). We are working on energy-efficient accelerators for BNN inference. We have worked to identify optimum entropy sources for RNG, then leverage novel devices to enable fully analog compute-in-memory, removing the need for digitizing said entropy. The resulting accelerator will be suitable for deployment on SWaPC-constrained systems and demonstrate improvements in BNN inference energy efficiency over state-of-the-art accelerators.  How emerging devices based on spin or ferroelectrics could be valuable in this space will also be discussed.