Publications

Reinforced Genetic Algorithm Learning for Optimizing Computation Graphs, ICLR 2020

Using Graph Neural Networks and Reinforcement Learning, we are able to propose good mutant sampling beta distributions to a genetic algorithm.

We use this genetic algorithm to optimize the placement and scheduling decisions of TensorFlow computation graphs in multi-device settings and discover that our algorithm can reduce the running time by upto 20% and peak RAM consumption by upto 50%.

Graph Reresentations for Higher Order Logic and Theorem Proving, AAAI 2020

We show that representing s-expressions as graphs and encoding them with Graph Neural Networks achieves state-of-the-art performance for automated theorem proving on the HOList benchmark.

There are several non-trivial ways to represent an s-expression as a graph and some of them are clearly better than others.

Zero-Shot Learning for Fast Optimization of Computation Graphs, NeurIPS 2019

A short version of our ICLR paper, I presented this at the ML for Systems Workshop at Vancouver in December 2019.