Neuralizing Algorithms

Investigator: Lee Wee Sun.

Most AI problems are computationally intractable in the worst case. The typical or average case could be much easier but is difficult to design for. We propose to use machine learning to learn algorithms with good average-case performance instead.

Specifically, we learn effective approximate inference algorithms for probabilistic graphical models using deep learning. The Factor Graph Neural Network can represent operations of a well-known approximate inference algorithm, the loopy belief propagation, and can potentially represent more powerful algorithms. It is learned from data, optimizing average-case performance on the data it is learned on, instead of being designed to work on all graphical models.