Exploiting Structure Information in Machine Learning

Machine learning has recently witnessed revolutionary success in a wide spectrum of domains. Most of these applications involve learning with complex inputs and/or outputs, which could be sequences and graphs in chemical and material design, functions and distributions in graphical models learning, and even control dynamics in reinforcement learning. The success of these applications of machine learning techniques often requires at least two factors: i) the exploitation of structure information in learning models, and ii) the utilization of huge amount of data. However, the structure information corresponds delicate conditions in optimization point of view, while a huge amount of data requires algorithms efficient and scalable. Integrating both parts can be very challenging, from both computational and theoretical perspectives.

In this talk, I share my research efforts on developing the principled, scalable and practical algorithms for large-scale structured learning. Specifically, I will discuss our reinforcement learning algorithm which exploits the recursion structure in Bellman equation. This work takes a substantial step towards solving the decades-long open problem in reinforcement learning for seeking a convergent algorithm with function approximatiors on off-policy data. I will also present our work on handling graphs as inputs in supervised learning. We proposed ‘structure2vec’ mimicking the mean-field inference as a general model for such tasks. Empirical results show the structure2vec achieves the state-of-the-art accuracy with smaller model size while faster training speed.

See more at https://www.microsoft.com/en-us/research/video/exploiting-structure-information-machine-learning/