Understanding and Improving Deep Generative Models

Lead Research Organisation: University of Oxford
Department Name: Autonom Intelligent Machines & Syst CDT

Abstract

Context & Potential Impact. The meteoric rise of the quantity of available data in the 21st century demands algorithms that can leverage this data without requiring expensive labels.
Deep Generative Models (DGMs) uncover rich patterns hidden within the data by learning to generate the data. These models are used to perform several tasks including: compression; clustering; representation learning; and density estimation. Improvements to these models could be deployed across a number of real world applications.
Aims and Objectives. In reality, DGMs have limitations. For example, training an automated decision making system using the representation learned by a DGM often yields unfair decisions. Further, DGMs can struggle to distinguish whether a novel input was drawn from the training distribution or not, leading to robustness concerns. This research aims to first understand the limitations of DGMs and then propose novel techniques to remedy these issues.
Research Methodology Novelty. This research will propose novel solutions to well known but often poorly understood limitations of DGMs.
Alignment to EPSRC Strategy and Research. This research fits within the Artificial Intelligence Technologies , Image and Vision Computing , Digital Signal Processing and Statistics and Applied Probability EPSRC research areas.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S024050/1 01/10/2019 31/03/2028
2243855 Studentship EP/S024050/1 01/10/2019 30/09/2023 Mrinank Sharma