Research Project Title:
Factor Graph for Reinforcement Learning at Scale
abstract:Reinforcement and deep learning have become popular in recent years, yet computing them at scale remains a problem. Popular scalable frameworks like MapReduce are not Turing-complete; developers who want to deploy learning algorithms at scale have to write their own data pipelines. The goal of this project is to build a hardware-agnostic Turing-complete computation framework based on the inference data structure of "factor graphs". Our factor graph language is Turing-complete and has been shown to perform better than state-of-the-art frameworks on single-machine architecture. We plan to continue building interfaces to popular hardware setups such as GPU clusters, so developers can focus on developing custom learning algorithms without being bogged down by infrastructure work.
I'm participating in SuperUROP because I want to gain a quintessential understanding of reinforcement, deep, and machine learning algorithms as well as obtain experience implementing these algorithms in single-machine and at-scale settings. I've taken machine learning courses at MIT and have been working on this project since early 2018. I hope to publish an open-source framework people would love to use.