Learning from Representations in Reinforcement and Unsupervised Learning

Lead Research Organisation: University of Oxford
Department Name: Statistics

Abstract

This project falls within the EPSRC Mathematical sciences research area.
This project lies in the context of reinforcement and unsupervised learning.
The number of data generated every day is steadily increasing. Analyzing and understanding this large volume of information is important. Indeed, creating algorithms so that computers can learn from the data they are given without any human intervention can give faster and more accurate results. This can have a lot of applications in our society and impact many areas socially and economically such as healthcare, transports, security ... For instance, in medicine, intelligent systems can assist doctors in their diagnosis through the analysis of radiographs. Another example is the analysis of images that can help policemen arrest individuals wanted. Consequently, developing both theoretical and computational methods and training systems to process the quantity and diversity of data we have is thus a topical and important challenge to be solved. So far, some systems have been successfully experimented to perform some very specific tasks. However, the common sense of intelligence, that is the perception and understanding of the world in which we are living in and that a child generally acquires by the age of two, has not been solved yet.
Unsupervised Learning shows promise to tackle this problem as this type of algorithms learn directly from data, without any pre-assigned labels. This project will focus on how leveraging the structure of the data and their representations can make learning more efficient.
Reinforcement Learning (RL) is concerned with how agents can learn behaviours by interacting with their environment. The applications of RL range from robotics, recommendation systems, aviation to healthcare. In the real world, agents often have to deal with large or continuous state spaces encoded in an unstructured way. Because we cannot represent the value of each state, we need to learn a structured representation from which we can express RL functions of interest in a more meaningful way. However, little is understood about what the right representations for a particular target task are. I aim at developing both theoretical and computational methods to provide answers to this question.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/N509711/1 30/09/2016 29/09/2021
2207522 Studentship EP/N509711/1 30/09/2018 29/09/2022 Charline Le Lan
EP/R513295/1 30/09/2018 29/09/2023
2207522 Studentship EP/R513295/1 30/09/2018 29/09/2022 Charline Le Lan