Key information:
| Student | Nikos Chazaridis |
| Academic Supervisors | Christine Evers, Tim Norman, Mohammad Belal |
| Cohort | 3 |
| Pure Link | Active Project |
Dissecting a complex notion into several ideas of higher granularity that are easy to process is an intuitive mental process that humans perform when trying to learn or teach others. Disentangled representation learning seeks to emulate this cognitive operation in artificial systems by enabling them to disentangle input signals into representations whose components refer to unique semantically meaningful concepts. The benefits of successfully generating disentangled representations are improving the performance of machine learning algorithms, providing explainability and fairness in algorithmic decision making and reducing model size as well as model complexity among others.
The objective of this research is to tackle disentanglement with a probabilistic perspective in time-series data thus leveraging temporal information, contrary to existing literature. We aim to design models that exploit temporal information and isolate sources of interest in data sequences based on Variational Inference and Autoregression.
This project is part of a research collaboration between the University of Southampton and the National Oceanography Centre, thus insights from disentangled representation learning will be applied to the problem of isolating events of interest in the marine environment. This poses a significant challenge due to two main reasons. Firstly, signals in the marine environment are typically a blend of various natural and anthropogenic sources making data labelling impractical and secondly, the sheer volume of data generated from monitoring the marine environment incurs substantial computational costs. We believe that disentangled representation learning and self-supervised machine learning may offer effective strategies for tackling this intricate task.
Contact: N.Chazaridis@soton.ac.uk