Size (OOD) Generalization of Neural Models via Algorithmic Alignment

Explore how neural models can achieve size generalization for algorithmic tasks through a conference talk that examines the challenge of enabling bounded-complexity neural networks to handle problem instances of arbitrary size. Learn about algorithmic alignment as a solution to this out-of-distribution (OOD) generalization problem, with theoretical foundations demonstrating how the combination of algorithmic alignment, sparsity regularization, and carefully selected training data enables provable size generalization when approximating the Bellman-Ford algorithm on arbitrary graphs. Discover practical applications through examples of designing efficient neural models for geometric optimization problems using algorithmic alignment principles, presented by Yusu Wang from UCSD as part of the Graph Learning Meets Theoretical Computer Science series at the Simons Institute.