Generative Modelling for Supervised, Unsupervised and Private Learning
Lead Research Organisation:
University of Oxford
Department Name: Engineering Science
Abstract
This project falls within the EPSRC AI and robotics theme. This project aims to understand how generative modelling approaches (specifically Generative Adversarial Networks - GANs) could be used to tackle several different problems: treatment effect estimation, feature selection, transfer learning, and private synthetic data generation. Specifically, we will investigate the question of whether the adversarial mechanisms that drive GANs can be adapted to other problem settings to define new loss functions suitable for solving problems slightly adjacent to the broader, more simple problem of pure generation.
The project will aim to identify which problems admit a form that can be solved using a GAN-type approach, primarily with the hope of identifying problems that can be grouped together (despite seemingly being unrelated) and addressed using the same modifications to the GAN framework.
The project, specifically the parts of it focusing on synthetic data and treatment effect estimation have a very big potential for real-world impact. The problem of estimating the effect of a treatment from observational data (collected "naturally" from the real-world, rather than through clinical trials, for example) has the potential to drastically improve the assignment of treatments in medicine, but also in many other fields where an intervention is taken, such as with job training programs, recommender systems. Synthetic data is a relatively new field with rapidly growing interest forming around it which has the potential to enable machine learning research at a much larger scale than is currently possible, shifting the power from those with data to those with machine learning skills by enabling data holders to share potentially sensitive data through the use of synthetic data, allowing the machine learning community at large to attempt to solve problems that otherwise do not present themselves in existing machine learning benchmark datasets.
The project will aim to identify which problems admit a form that can be solved using a GAN-type approach, primarily with the hope of identifying problems that can be grouped together (despite seemingly being unrelated) and addressed using the same modifications to the GAN framework.
The project, specifically the parts of it focusing on synthetic data and treatment effect estimation have a very big potential for real-world impact. The problem of estimating the effect of a treatment from observational data (collected "naturally" from the real-world, rather than through clinical trials, for example) has the potential to drastically improve the assignment of treatments in medicine, but also in many other fields where an intervention is taken, such as with job training programs, recommender systems. Synthetic data is a relatively new field with rapidly growing interest forming around it which has the potential to enable machine learning research at a much larger scale than is currently possible, shifting the power from those with data to those with machine learning skills by enabling data holders to share potentially sensitive data through the use of synthetic data, allowing the machine learning community at large to attempt to solve problems that otherwise do not present themselves in existing machine learning benchmark datasets.
Organisations
People |
ORCID iD |
James Jordon (Student) |
Studentship Projects
Project Reference | Relationship | Related To | Start | End | Student Name |
---|---|---|---|---|---|
EP/N509711/1 | 30/09/2016 | 29/09/2021 | |||
2618232 | Studentship | EP/N509711/1 | 30/09/2018 | 31/12/2021 | James Jordon |
EP/R513295/1 | 30/09/2018 | 29/09/2023 | |||
2618232 | Studentship | EP/R513295/1 | 30/09/2018 | 31/12/2021 | James Jordon |