Keynote III: Boolean Logic Networks for Machine Learning

Summary form only given, as follows. The complete presentation was not made available for publication as part of the conference proceedings. This talk explores the use of logic networks and Boolean methods in machine learning (ML). Both synthesis and verification are addressed. On the synthesis side, we introduce a novel ML model based on logic networks, which can potentially replace neural networks in some applications. The advantages are, substantially reduced evaluation latency, simpler hardware implementation (no need for memories and arithmetic operations), straight-forward design automation. The challenge is, matching the accuracy of neural networks. On the verification side, we use logic networks, in particular, and-inverter graphs (AIGs), to detect overfitting in any ML model using only the AIG representation of the model and the training data, assuming that the evaluation data is not available or cannot be trusted. BIOGRAPHY