Activation-Kernel Extraction through Machine Learning

Machine Learning flooded many research fields, including Electronic Design Automation (EDA). The availability of algorithms that can solve complex problems through generic rule formulations represent a fresh opportunity to improve existing design paradigms. In this work we investigate the use of machine learning to manipulate logic circuits. More specifically, we envision the use of Classification and Regression Trees (CARTs) as tools for modeling generic Boolean functions through a representative subset of core expressions, what we call the activation kernels. Experiments conducted on a subset of open-source benchmarks demonstrate that CARTs are indeed able to identify the activation kernels that cover whole Boolean functions with a high degree of accuracy (89% on average). In order to quantify other figures of merit, we also provide a physical implementation of such activation kernels. Results show that the obtained circuits are amazingly smaller than standard-cell based circuits synthesized through a classical logic synthesis flow (16X less devices on average).