Programming with Neural Surrogates of Programs

Explore the concept of neural surrogates in programming through this insightful talk by Alex Renda from MIT CSAIL. Delve into three case studies of surrogate-based design patterns, evaluated using a large-scale CPU simulator. Learn about two innovative approaches for constructing optimal program surrogates: the Renamer, a Transformer architecture invariant to semantics-preserving variable renaming, and Turaco, a methodology for sampling datasets to train neural network-based surrogates. Gain valuable insights into programming methodologies in the era of deep learning and discover how surrogates can enhance the speed and accuracy of complex programs.