Human-in-the-loop AutoML for the Automation of Data Science and Artificial Intelligence, Computer Science, Applied Mathematics, Electric Engineering - PhD (Funded)

Project Description:

Data science (DS) and artificial intelligence (AI) have achieved considerable successes in recent years and an ever-growing number of disciplines rely on it. Building effective machine learning (ML) models from data is an integral part of the majority of DS and AI workflows. The complexity of these tasks highly rely on human ML experts thus are often beyond non-ML-experts. The rapid growth of ML applications has created a demand for off-the-shelf ML approaches that can be used easily and without expert knowledge. This is also known as automatic machine learning (AutoML) which automatically obtain a ML pipeline from data and have become an increasingly popular area in the ML community.

In practice, the design of an AutoML system is essentially a multi-objective optimization problem with more than one performance criterion to consider simultaneously. These criteria, such as predictive performance, model size, prediction speed, fairness and interpretability will vary between projects and have inherent trade-offs. ML pipeline is design by human thus should work for human. Unfortunately, only a few efforts have focused on the interaction between human and AutoML. Partially due to this concern, the current AutoML is still another black box for both ML experts and non-ML-experts to use in practice.

In this project, we aim to develop a revolutionary AutoML system that keeps human-in-the-loop. It consists of the following objectives.

1. Develop interactive data-driven evolutionary multi-objective optimization approaches that proactively learn the decision makers’ preference and progressively guide the search towards the solution(s) of interest.
2. Develop a human computer interaction (HCI) platform (not limited to high-dimensional visualization techniques) that plays as the interface to bridge the decision maker and the computer.
3. Develop effective preference learning approaches to model human preference and reward/fitness function(s) from multimodal data. In addition, we are also interested to investigate and understand the human’s role and mental behavior during the interaction from the perspective of eye tracking, EEG or even fMRI data, which can feedback into a better HCI design.
4. Develop inverse modeling techniques to unbox the black box. This provides interpretability to the trade-off solutions and multi-criterion decision-making. In particular, we are interested in modeling and understanding the dependency between decision variables and multiple conflicting objective functions, as well as the inverse mapping between objective functions to decision variables by using neural symbolic regression.

Note that the successful PhD student is not expected to carry out all these four objectives in this project but just selected ones pertinent to her/his research interests and experience.

This award provides annual funding to cover Home tuition fees and a tax-free stipend.  For students who pay Home tuition fees the award will cover the tuition fees in full, plus at least £17,668 per year tax-free stipend.  Students who pay international tuition fees are eligible to apply, but should note that the award will only provide payment for part of the international tuition fee and no stipend.  

The studentship will be awarded on the basis of merit for 3.5 years of full-time study to commence on 25th September 2023. The collaboration with the named project partner is subject to contract.  Please note full details of the project partner’s contribution and involvement with the project is still to be confirmed and may change during the course of contract negotiations

Advert information