Multi-task Learning in Deep Neural Networks

In AI and Deep Learning a fascinating challenge is to create an agent that can solve multiple tasks. In the large majority of machine learning approaches, the trained model is specialised on a single task and is useless to solve any other problem. Multi-task Learning (MTL) is an approach to machine learning in which a model is trained to solve multiple tasks simultaneously. This is similar to the learning process in humans, who learn general skills useful to multiple tasks: e.g., hand dexterity is useful to solve many tasks and is improved by learning many tasks at the same time. In machine learning MTL has been shown to be particularly effective in generating better generalised models that take advantage of the similarities and differences across tasks. Its effectiveness is also very useful in solving a group of problems altogether, each of which with a limited number of training records. In this PhD project MTL methods will be investigated to identify the best approaches for Deep Neural Networks and, in particular, for training the networks with evolutionary algorithms instead of a more classic gradient descent strategy. The overarching aim of this project is to contribute to theory and applications of machine learning with effective deep Neuroevolution algorithms and non-linear optimization methods. A key enabling factor to understand deep neural networks is to consider neural networks as complex networks, this PhD project will try to decipher the complex neural networks by the information theoretic learning principle, in general, and the information bottleneck, in particular. Multi-task Deep Neuroevolution will be applied to classic and recent testbeds for deep learning (e.g., CIFAR-100, Atari games, XTREME, StarCraftII) as well as real-world problems, such as the prediction of neurodegenerative diseases (dementia) from human brain images or synthetic motor control for modelling rich and diverse motor behaviours across multiple tasks at humanoid scale. The project will be guided by a team of experienced supervisors with extensive competence in this area and will have access to state-of-the-art computing facilities.

For further information email to [Email Address Removed].

To apply for this studentship please submit an application for a PhD in Computer Science at http://www.reading.ac.uk/graduateschool/prospectivestudents/gs-how-to-apply.aspx

1) Please quote the reference ‘GS20-097’ in the ‘Scholarships applied for’ box which appears within the Funding Section of your on-line application.

2) When you are prompted to upload a research proposal, please submit a personal statement (max 1 page) to present your interest, motivation and suitability for this project.