Efficient and Transparent Machine Learning: Exploring Applications of Differentiable Logic Gate Networks

Abstract

Despite their remarkable capabilities, the decision-making process of deep neural networks remains a black box, obscuring the logic behind model decisions and undermining trust in critical domains such as healthcare, national security, and law. While the field of Explainable Artificial Intelligence (XAI) has emerged to mitigate these issues, current post hoc explanation methods often fail to deliver reliable insights as they are applied after a model has been trained, rather than addressing the underlying architecture. To overcome these fundamental limitations, this thesis explores Differentiable Logic Gate Networks (DiffLogic), a type of neurosymbolic AI (NSAI) architecture that enables neural networks to learn a distribution of logic gates for each node. Although DiffLogic was originally introduced with a focus on accelerating inference speeds, its potential for explainability has remained largely unexplored. Specifically, this work seeks to address these gaps by: (1) implementing DiffLogic on Field Programmable Gate Arrays (FPGA) for hardware acceleration, (2) developing a compression algorithm for DiffLogic models, and (3) deploying visualization tools to intuitively interpret learned logical structures.