AutoPC: An Open-Source Framework for Efficient Probabilistic Reasoning on FPGA Hardware

In the quest for more advanced and energy-efficient edge AI, probabilistic reasoning models can complement or replace deep learning (DL) models, as they are generative, explainable, and trustworthy. However, their hardware implementation and acceleration are still in the early stages compared to DL due to more ad-hoc implementations and challenges translating them into computational steps. This recently evolved with Probabilistic Circuits (PCs), which can be trained with mainstream software and lead to more hardware-efficient inference. Yet, there is currently no single open-source framework dedicated to computing PCs on hardware. In this work, we introduce such a framework called AutoPC, allowing us to (1) compare PCs trained with different PC algorithms to find the most suited, (2) find the optimal resolution required for hardware computation with minimal cost, and (3) automatically generate FPGA hardware for executing PC models with high speed (40–200 GOPS) up to the FPGA capacity. We hope AutoPC serves as a baseline to showcase the possibilities of probabilistic reasoning and broaden the use of PCs.