Abstract Interpretation for Automatic Differentiation

Explore the application of abstract interpretation to automatic differentiation (AD) in this 11-minute conference talk from ACM SIGPLAN's LAFI'24. Delve into the presenters' argument for the necessity of precise, general, and scalable abstract interpretation techniques specifically tailored to AD code. Discover how leveraging the inherent structure of AD can enhance analysis capabilities. Learn about a novel use case demonstrating the power of abstract interpretation in analyzing Coordinate MLPs from Graphics, showcasing significant improvements in precision compared to previous approaches. Gain insights into the potential for advancing static analysis of AD code, which is crucial for machine learning, scientific computing, and graphics applications.

[LAFI'24] Abstract Interpretation for Automatic Differentiation