Towards Reverse Algorithmic Engineering of Neural Networks

Explore a lecture on reverse algorithmic engineering of neural networks that delves into explaining complex machine learning models from an algorithmic perspective. Learn how the NeuroSAT algorithm for SAT solving demonstrates the discovery of learned combinatorial features and algorithmic concepts in neural networks. Understand how machine learning models can outperform traditional algorithms in combinatorial optimization tasks while maintaining explainability. Examine a framework that reveals how NeuroSAT learns general algorithmic concepts similar to statistical inference methods, including computing confidence levels, fixing variables with highest confidence, and solving residual formulas. The presentation, delivered by Dr. Dan Vilenchik from Ben-Gurion University, draws from collaborative research with students from BGU, Tel-Aviv-Yafo Academic College, and MIT, offering insights into making complex neural networks more interpretable and trustworthy.