Learning disentangled representations to achieve responsible and transparent AI

Learning representations of input data is a subtask of any modern machine learning model: we learn a useful data representation that is helpful to achieve our main goal whether it is to classify an image or to learn an agent to play a game.

Disentangled representations are aimed to model different factors of data in different dimensions. For example, representations can be learnt to be disentangled into class-related and class-independent components, such that if we want to build a model to generate cat and dog images class-related components would determine ear and nose shapes and class-independent components would govern fur and eye colours.

It has been shown that disentangled representations can help with overall performance increase and in particular, for example, they may help with diversity of generated images in image-to-image translation task [1], or help with privacy preservation in voice assistant systems [2]. This project will research further how disentangled representations can help to build responsible and transparent AI systems.

One of the directions here can be tackling the problem of diversity and bias in AI. The example above shows that disentangled representations lead to increased diversity in generative modelling, which in turn can be used to transfer learning and domain adaptations. In this case diversity of generated content will increase performance of a downstream task in a new domain.

In addition to increasing diversity, disentangled representations can also be used to tackle hidden biases in data. It was shown that it is not enough to remove direct indicators of protected characteristics in training data for a machine learning algorithm. These characteristics can be inferred from the remaining data, such as in the simplest case a gender can be often be inferred from a name. Disentangled representations can be used to factorise components of learnt representations into protected characteristic-specific and independent thus providing a tool to truly discard all the information related to protected characteristics and to ensure an unbiased algorithm for a downstream task. 

The above methods would assist transparency of machine learning systems already but it can be improved further with learning interpretable disentangled representations. In this case we do not only factorise learnt representation but make it explicitly interpretable, e.g., the first component would be responsible for hair colour, the second component would be responsible for hair shape, etc. 

This research project will be based in the UKRI Centre for Doctoral Training in Accountable, Responsible and Transparent AI (ART-AI). We value people from different life experiences with a passion for research. The CDT's mission is to graduate diverse specialists with perspectives who can go out in the world and make a difference.

Informal enquiries about the project should be directed to Dr Isupova (oi260@bath.ac.uk).

Applicants should hold, or expect to receive, a degree in Computer Science, Mathematics, Statistics or related areas. A strong mathematical background and programming experience are desirable.

Formal applications should be accompanied by a research proposal and made via the University of Bath’s online application form. Enquiries about the application process should be sent to art-ai-applications@bath.ac.uk.

Start date: 2 October 2023.