CSC (China Scholarship Council) PhD project proposals for 2019

Title: Improving the manufacture of composite components

Carbon fibres were first created in 1860 for use in light bulbs, but the first significant use of carbon fibre in structural materials took place nearly 100 years later. Although carbon fibre has been used a composite material for nearly 50 years, it is still true that the manufacture of composite structures is still as much art as science. This project, in collaboration with the National Composites Centre will set out to understand the manufacturing process better and improve the quality of the finished product.

The project will apply statistical and machine learning techniques to analyse each stage of creating a component: a unique dataset is being created with detailed measurements of many characteristics of the fabric and the manufacturing processes. This includes visual inspection, images of fabric layers, thickness and airflow, the resin infusion parameters, image analysis of the surface, and finally a group of test measurements including corner strength, microscopy, glass transition temperature, and destructive testing.

This project will suit a student who wants to develop new machine learning techniques for a challenging real-world problem.

Contact: Ian Nabney


Title: Enhanced data visualisation for large datasets

While a range of projective methods exist to facilitate users obtaining an overview of the data they are analysing, these do not scale well to Big Data, require expert knowledge to utilise and are often very difficult to interpret in practice. It is also important that users are confident that the projective visualisations provide a faithful representation of the underlying data.

We will develop methods for visualising heterogeneous Big Data with minimal input from the user by employing modern machine learning inference methods, such as variational approaches which replace an inference problem with an optimisation problem. This will focus initially on the Generative Topographic Mapping (GTM) method, and consider hierarchical methods for massive data sets. We will also research suitable methods for providing users with clear evidence of the ëgoodnessí of a non-linear projective visualisation. In particular this will present the user with some metrics summarising the ëfití of the visualisation to the underlying data using the Kullback-Leibler distance for probabilistic methods such as GTM.

Having constructed a projective visualisation of complex data it is critical the user can understand this. Current interactive visualisation methods are powerful but are limited by their reproducibility due to the human judgements involved. This limits their utility in decision making, since communicating the route by which the final interpretation was reach through data analysis is critical to others trusting the outcomes of the work. We will develop a novel system to capture the userís navigation through the data, choices made in data cleansing, modelling, and viewing. This will enable the users to annotate and capture their interactive visualisation to share it with colleagues to improve cross-team understanding.

Contact: Ian Nabney


Title: AI-driven Decision Support for Financial Modelling

The tendency of combining financial studies with artificial intelligence technologies has attained significant scientific attention in recent years. Financial industries and domains in essence revolve around market data and human behaviors, which can be understood and analyzed by machines with the integration of suitable interactive mechanisms. Historical studies on financial problems like portfolio management, risk allocation, market forecasting and credit rating have already been investigated by the AI technology community, with particular emphasis on highly efficient and objective solutions. While these interdisciplinary studies have shown great potentials, the water has not been fully taped. Under more sophisticated scenarios where (i) both financial information stemming from disparate formats and sources, and (ii) diverse participants with different distinctions coexist, it will be unrealistic for human beings alone to fully analyze the situation and make effective decisions. A deliberately designed interactive AI system based on data mining, machine learning and decision-making techniques with human analysts in the loop would provide us with the auxiliary we need.

This PhD project concentrates, on one hand, on devising an interactive intelligent system that can analyze, process and cluster real-time data into patterns predicated on knowledge extracted from historical financial data, especially under fuzzy or linguistic information environments. Time series of multiple data sets from financial market and market participants are used as training samples to help machines to develop a comprehensive understanding of different situations, and where a certain pattern is exhibited by data, the machine will be able to tell users what kind of situation they are in or about to going through. On the other hand, the influence of possible investment alternatives or market events will be investigated and benchmarked to support reliable decision making jointly made by human analysts and financial analytics systems. Different alternatives will be tested through the training system and the outputs will be in the same formats of former training samples, which if combined with multi-agent or multi-stage decision making models can be used to evaluate the decision alternative(s). In the end, we will be able to rank alternatives and come up with the best one. Applications of the proposed research can be rather broad like personal or group investment decision making, financial market fraud detection and financial distress forecasting. This study will be of great potential and application value to advance state-of-the-art AI-based approaches for financial analysis and decision-making.

Contact:Ivan Palomares Carrascosa


Title: Learning Interpretable Changes from High-dimensional Time-series Data

The change-points in time-series (such as stock prices) indicate potential phase-transition points of the underlying system and has an important value to data science. However, there has been increasing demands on analyzing time-series with very high dimensionality (such as Neuroimaging data or Gene Expression data) recently. Classic statistical modelling approaches fail terribly on these datasets. This project explore methodologies of analyzing and interpreting changes on these high-dimensional datasets using latest ideas from high-dimensional statistics.

We will explore various time-series tasks such as change-point (anomaly) detection, structural change detection and causal inference, together with domain experts in biology and neuronscience to answer real-world research questions. We will also work with statisticians at Math School to further advance these methodologies from theoretical point of view.

*This project requires students with solid mathematical statistics background.

Contact: Song Liu


Title: Beat Machine Learning by Machine Learning

As more and more Machine Learning algorithms have been integrated into various systems, the ìrobustnessî of machine learning has become an increasing concern for companies as some systems operate on an ìadversarial environmentî: Your opponents are very happy to see your ML system breaking down and they will try to make it happen. To build a robust system, we need to understand how these potential attacks may happen. This project tries to simulate such attacks on some common ML systems and investigate potential defence mechanisms.

One way to engineer such attacks, as you guessed, is to use ML itself! This project will explore several ways to attack or defend attacks under different settings (classification, regression, etc.) and understand the vulnerability of a ML system so future ML developers can build a more robust system.

Contact: Song Liu


Title: Opinion diffusion logics for swarms

A fundamental challenge in swarm robotics is that of how best to propagate information across the population on the basic of only local interactions between robots. In a complex noisy environment it may be more appropriate to represent knowledge qualitatively rather than quantitatively. For example, in comparing different possible decisions, actions or policies it may be sufficient to identify a partial ordering of these rather than attempting to quantify the value of each. This project will investigate possible approaches to opinion diffusion in robotic swarms in which individual agentsí beliefs take the form of logical formulas rather than being represented by numerical values. More specifically, we will study how such opinions propagate under a range of belief combination models. The project will involve a mixture of simulation and Kilobot experiments in which robots will hold qualitative beliefs in the form of canonical logical formula. There will also be an opportunity for theoretical studies using a combination of formal logic and network theory.

Contact: Jonathan Lawry


Title: The Art of Persuasion in Distributed AI

The study of how opinions evolve and converge in distributed systems in an important and
established research area with applications in social networks, swarm robotics and distributed AI. In most of this research, agents combine their opinions based on simple operators, e.g. weighted average, constrained only by the similarity between their beliefs or physical factors such as proximity. However, for humans to change their beliefs there is usually an element of persuasion involved in which one agent puts forward an argument to attempt to convince another to adapt their opinions. Persuasion is also likely to be important for explainable AI systems which need to justify their decisions and actions. In this project you will investigate agent-based models of opinion formation based on persuasion. The representational framework for expression will be probability
logic and arguments will be viewed as logical formulas which are believed (to some
degree/probability) by both agents in a dialogue. The aim of each agent will be to identify such a formula that will be accept by the other and to consequence more their opinion closer to their own.

Contact: Jonathan Lawry