Research
Neural algorithmic reasoning.
I study how neural networks learn and use algorithmic structure. Three common topics in my work are generalization on graphs, discrete and combinatorial optimization, and algorithmic alignment.
Generalization on graphs
Graphs are a natural setting for studying NAR because they offer a rich collection of simple but powerful algorithms, and many problems can be cast as graph problems. I study when graph neural networks learn these algorithms and generalize to larger or differently distributed graphs. Our Bellman–Ford GNN work establishes conditions under which size generalization is guaranteed for the shortest-path problem. Which Algorithms Can Graph Neural Networks Learn? develops a broader framework, characterizing a large class of algorithms for which learnability and size generalization can be guaranteed.
Discrete & combinatorial optimization
Combinatorial optimization is a particularly promising application of neural algorithmic reasoning. Classical algorithms still outperform neural combinatorial optimization in most settings, motivating neural models endowed with the strengths of traditional algorithms. Our Birkhoff extension provides differentiable objectives over permutations with rounding guarantees. NN-Steiner combines neural components with a classical approximation framework for the Steiner tree problem.
Algorithmic alignment
Algorithmic alignment is one of neural algorithmic reasoning’s most powerful techniques. By designing neural architectures that are structurally similar to a target algorithm or algorithmic framework, we can create an inductive bias toward algorithmic behavior. This approach promotes size generalization, efficient computation, and the use of a problem’s mathematical structure. NN-Steiner builds a neural model aligned with Arora’s approximation algorithm for Steiner trees. Our Bellman–Ford GNN work uses the alignment between message-passing neural networks and the Bellman–Ford algorithm as a case study to analyze when algorithmic alignment yields provable benefits.
Earlier work and broader perspective
My past work is in experimental physics, mathematical biology, and quantum algorithms. More broadly, I am interested in new paradigms in computing and how traditional algorithms can be combined with these paradigms to solve problems more effectively. I find applications such as biology, physics, and chip design essential for motivating research questions and understanding when these approaches are most useful.