Current PhD Projects

'Scalable Track Analytics'

Nowadays, shipping is one of the most important industries that undeniably requires increased security and safety. Both security and safety can be improved through maritime analytics applied on the massive amount of currently available heterogeneous maritime data. Using data from multiple sources, such as CCTV, AIS and RADAR, tasks such as vessel type classification, vessel behavioural analysis, and anomaly detection can effectively be addressed. In this project we focus on ship type classification using novel methods and scalable maritime analytics over AIS and RADAR streams.

Download here: Scalable Track Analytics 

'Coordination and Cooperation in Adversarial Engagements'

This project will develop methods to provide tactical guidance and decision making for future air defence systems. The aim will be to provide an operator that is robust, timely and that optimises the defensive responses across all available assets; including platform/vehicles, countermeasures and inceptors.

Download here: Coordination and Cooperation in Adversarial Engagements 

'Drone Detect and Desist'

Due to the increased use of drones for recreational and commercial purposes, safety measures need to be taken to ensure airspace and civilian security. Drone detection can solve these problems by applying video processing techniques and artificial intelligence. When creating such systems, there are many challenges to consider such as, harsh environmental conditions, obscure sceneries, long operating duration and low energy consumption. This research is fascinating because it involves interdisciplinary studies and is useful and practical.

Download here: Drone Detect and Desist 

'High Performance Processing of Distributed Acoustic Sensing Data: Turing Optical Fibers into Massively Parallel Microphone Arrays.'

The project is about the development of new detection and tracking algorithms to apply to optical fibre sensors. The idea is to turn Distributed Acoustic Sensor (DAS) into an array of microphones and to use the data provided to monitor highways, railways, and critical infrastructures, like oil pipelines. A new Bayesian signal process chain based on machine learning algorithms will improve the performance and decrease the complexity of the algorithm in a real-world scenario.

Download here: High Performance Processing of Distributed Acoustic Sensing Data 

'Uncertain Heterogeneous Algorithmic Teamwork'

This project is focused on showing that raw aggregated capability can outperform carefully constructed co-ordination. Specifically, there are situations where vast quantities of computational hardware in a heterogeneous environment can outperform dedicated high-performance resources in a carefully constructed homogeneous environment. This PhD project will investigate whether it is possible to adapt a divide-and-conquer algorithm (which is at the heart of set numerical Bayesian techniques) to operate efficiently in an uncertain heterogeneous environment.  The aim is to develop the infrastructure that makes it possible for vast heterogeneous computing resources (e.g. desktop PCs, GPUs and android phones) to operate effectively in a team.

Download here: Uncertain Heterogeneous Algorithmic

'Distributed Exploration and Exploitation with Passive RF Sensors'

This PhD project is about cooperative navigation of unmanned air vehicles (UAVs) and exploration of an area using passive RF (radio-frequency) sensors under various constraints and reachability objectives. Passive RF sensors pick up the signal that an object emits and provide information about its angular position and signal strength. There are many challenges in this project depending on its setting, for example, in a search and rescue scenario, under hazardous weather conditions, the goal could be to search an area quickly, by avoiding damaging weather and at the same time maximising the probability of finding survivors. We can identify the following two questions that are of interest here, "how fast can several UAVs complete the search of multiple objects with a mathematical guarantee?" or "is there an efficient protocol to guarantee safe and reliable communications under several constraints?".

Download here: Distributed Exploration and Exploitation with Passive RF sensors

'Fast Bayesian Deep Learning'

Deep learning has become one of the fields of Artificial Intelligence research that raises great interest in machine learning researchers due to its very high performance in specific tasks. However, deep learning encounters difficulties in many fields due to its absolute output certainty which might be detrimental for decision-making tasks. The goal is to couple Bayesian methods with deep learning to build complex models with uncertainty measurements. The current implements of Bayesian neural networks are accurate OR fast, the objective now is to meet both these aspects using new sampling methods.

'Fast Fully Bayesian Gaussian Processes'

This project is focussed on adapting Gaussian Processes for the age of Big Data, not only in terms of scalability of the method but also in limiting the need for non-autonomous interaction. The need to specify both mean and covariance functions, which then need to be tuned such that they best represent the training data, limits a Gaussian Process' ability to learn from the data in an unsupervised manner. Recently, there has been a resurgence in research into scalable approaches for Gaussian Process regression from the wider community, both in improvements to existing algorithms and in exploitation of modern computer architecture. The scope for improving the scalability of GPs further while also integrating the possible ideas of an autonomous statistician gives way for this project.

Download here: Fast Fully Bayesian Gaussian Processes 

'Scheduling Surveillance of Space Objects'

The aim of this project is to develop cutting edge sensor surveillance strategies to observe space objects which can outperform existing techniques by developing high quality, efficient, non-myopic sensor management algorithms to control space-surveillance sensors. These algorithms are required to maximise the value gained from limited sensor resources to enhance the understanding of the composition of space objects in orbit, specifically using ground-based optical telescopes. These surveillance strategies should have the ability to detect and respond in the instances that a satellite manoeuvre was performed and/or a collision event has occurred, to ensure that the correct data is promptly collected to enable mitigation actions.

Download here: Scheduling Surveillance of Space Objects

'Data-driven Intelligence for Countering Crime'

Organised crime is one of the greatest threats to the UK’s national security. The role of the National Crime Agency (NCA) is to protect the public by disrupting and bringing to justice those serious and organised criminals who present the highest risk to the UK. However, the volumes of data are growing exponentially. There is therefore a need to work with these experts to automate the processing of the data, learn what these hallmarks are and then search the data for instances worthy of further human analysis. The need is growing for such data-driven intelligence.

Download here: Data Driven Intelligence for Countering Crime

'Learning to See More: Better Bayesian Track Before Detect using Statistical Machine Learning'

This project is concerned with improving the ability to detect faint objects in, for example, radar data. By improving detection performance, cheaper sensors will be able to emulate more expensive sensors, making the development of advanced detection algorithms very important in industrial settings. Leonardo develops such sensors.  The key challenges are in developing real time processing methods for distributed processors that can use low-power processor systems and using adaptive scheduling to maintain energy efficiency across a number of processors.

Download here: Learning to See More Better Bayesian Track Before Detect 

'Machine Learning to Identify Unique Events in Sparse Hyperspectral Datasets'

The goal of this PhD project is to investigate machine learning approaches that may permit images from stable materials, obtained from a wide variety of methods, to be used to increase the ability of methods such as hyperspectral imaging in electron microscopy and X-ray systems to observe thermodynamically unstable materials and processes on the atomic scale. Such advances have the potential to significantly impact the search for new personalized medicines, the development of new advanced energy storage systems, and our ability to directly see chemistry important for catalysing environmentally friendly processes.

Download here: Machine Learning to Identify Unique Events in Sparse Hyperspectral Datasets 

'Distributed Hypothesis Generation and Evaluation'

Approaches should be developed to allow analysts with particular expertise to contribute their knowledge to different parts of an intelligence analysis in a collaborative manner by generating sub-arguments which are coherent with the analysis as a whole. Analysts or agents, using a well-defined theory of abstract argumentation, may then evaluate the hypotheses, which may suggest further collection of information or require refinement of hypotheses. Intelligence analysis should be considered as a cycle and therefore all stages of the process should be compatible with collaborative and distributed analysis. The approaches developed should be validated against other intelligence analysis techniques. Care will be needed to mitigate the potential for biases in the system.

Download here: Distributed Hypothesis Generation and Evaluation'

'Towards Data Driven Aerodynamic Models: Data Fusion of Experiment and Simulation'

The digital age with ubiquitous physics-based computational engineering tools, such as computational fluid dynamics (CFD), machine learning algorithms and ever-increasing computing power, helped accelerate the development of novel technologies deployed in the civil transport sector, as well as in defence and security, to meet the most demanding economic, environmental and societal challenges.  It is envisaged to first explore future algorithms, including AI surrogate models, for near real-time joint experimental/numerical data analysis, that is uncertainty-aware, robust and quantifiable, to inform and optimise a wind-tunnel campaign, including on-the-fly. Second, considering the vast amount of data that a high-fidelity CFD run and a fully instrumented wind-tunnel test can produce, particularly for unsteady flow simulations, the first objective calls for high parallelisation utilising future computing systems, such as those explored within this CDT.

Download here: Towards Data Driven Aerodynamic Models 

'Extracting Important Information from Noisy Spectra'

The aim of this project is to develop algorithms that are both fast and accurate. The proposed approach is to speed up mature numerical Bayesian algorithms that are already accurate. These algorithms are accurate because they search over the space of all combinations of peaks that could be present in each spectrum. The algorithms explicitly compare the measured spectrum with those associated with each combination and also explicitly consider the extent to which the combination is consistent with prior knowledge (Gencoa are developing libraries to make it possible to predict which set of peaks would be likely to co-occur; a single impurity will typically give rise to multiple peaks). This approach makes it possible for the algorithms to make accurate inferences about whether a bump in the noise-floor or on the side of a large peak is caused by chance or by the presence of a low-amplitude peak.

Download here: Extracting Important Information from Noisy Spectra

'Data-driven Intelligence for Countering Crime'

Organised crime is one of the greatest threats to the UK’s national security. The role of the National Crime Agency (NCA) is to protect the public by disrupting and bringing to justice those serious and organised criminals who present the highest risk to the UK. However, the volumes of data are growing exponentially. There is therefore a need to work with these experts to automate the processing of the data, learn what these hallmarks are and then search the data for instances worthy of further human analysis. The need is growing for such data-driven intelligence.

Download here: Data Driven Intelligence for Countering Crime 

Chris Blackman     

‘Machine Learning for Target Detection and Classification Using Multi-modal Airborne Sensor Data’

This PhD will seek to establish whether multi-modal airborne sensor data, used early in the signal processing chain, can be used to detect and classify objects earlier in the processing chain leading to improved detection and classification performance.

Download here: Machine Learning for Target Detection and Classification 

‘Scheduling of Distributed Information Processing’

The focus of the PhD will centre on developing statistical emulators that can predict how computation will process data and generate information. Once those emulators exist, the focus will be on using the emulators to schedule distributed computational resources.

‘Bayesian Reinforcement Learning for Control of Continuous Industrial Processes’

This project is focussed on finding a good control strategy to optimize the quality of a glass product during a specific manufacturing process or reducing the cost of this process through less faulty goods. Ideally, both goals may be combined. The widely used Proportional-Integral-Differential (PID) controller should thus be replaced or expanded at best. Although the PID controller is widely used, it has its weaknesses and sometimes requires human intervention. To overcome them Bayesian Inference will be used to generate simulated data based on offline historic data from the manufacturing process.

Download here: Bayesian Reinforcement Learning for Control 

Jinhao Gu

‘Artificial Intelligence for Fast Discovery of Novel Materials for Healthcare’

This PhD will incorporate the following elements: (1) Review of existing approaches to using Gaussian Processes to solve regression problems involving descriptions of molecules as the features and existing techniques for articulating similarities between different molecules and/or sets of molecules; (2) Development of scalable Gaussian Process implementations (e.g. involving variational or distributed approaches to representing the uncertainty) and hyper-parameter estimation (e.g. using novel Sequential Monte Carlo methods) that exploit emerging many-cored compute resources to facilitate timely performance; (3) Application of these approaches in the context of Bayesian Optimisation to help answer questions pertinent to the discovery of novel materials for healthcare.

Download here: Artificial Intelligence for Fast Discovery of Novel Materials for Healthcare

 William Jeffcott

‘Data Science and Artificial Intelligence for Smart Sustainable Plastic Packaging’

High-density polyethylene (HDPE) is widely used in plastic production. It can be recycled to produce a post-consumer resin (PCR), which can then be used to make new plastic products. However, this PCR may contain differing grades of plastic, and may become contaminated with other materials. These factors affect the performance of the end product. At present, the more favourable solution for companies is to make new virgin plastic rather than recycling.

Download here: Data Science and Artificial Intelligence (Plastics) 

 George Jones

‘Non-Myopic Approaches to Sensing and Surveying’

A key focus of the project is likely to be understanding the potential for exploiting multi-core computing hardware such as GPUs and/or FPGAs. Another area of interest is how such approaches can generate solutions as part of a human-machine teaming concept, in order that they promote trust and transparency and help end-users understand the inherent trade-offs in the optimisation process.

Download here: Non-Myopic Approaches to Sensing and Surveying 

Andrew Millard

‘Scalable Online Machine Learning’

This project relates to extending the state-of-the-art to enable machine learning to fully capitalise on the information present in never-ending data streams. The additional data that arrives over time contains information that should facilitate improved machine learning. Not using this information gives rise to consistent yet surprising errors: this typically occurs when the training data is small relative to the algorithm’s empirical experience.

Download here: Scalable Online Machine Learning 

Kieron McCallan

‘Using Next Generation Machine Learning and AI techniques to Aggregate Information Pertinent to Automotive Glazing’

Simulation techniques exist to assess how feasible it is for existing production processes to be used to make a new glazing part for a car. While instances of these simulations have been run extensively in the past, it is unlikely that the new glazing part will be identical to one that has been assessed previously. 

Download here: Using Next Generation Machine Learning and AI Techniques

Joshua Murphy

‘Scalable Online Machine Learning’

The project relates to extending the state-of-the-art to enable machine learning to fully capitalise on the information present in never-ending data streams. The additional data that arrives over time contains information that should facilitate improved machine learning. Not using this information gives rise to consistent yet surprising errors: this typically occurs when the training data is small relative to the algorithm’s empirical experience.

Download here: Scalable Online Machine Learning

Jack Taylor 

‘Optimized Sampling Approaches for Compressive Sensing in Multi-Dimensional Datastreams’

The goal of this PhD project is to determine the minimal level of sub-sampling that will be sufficient to reconstruct images from transmission electron microscopes. By developing and implementing new compressive sensing algorithms to tackle the challenge of hyperspectral data, the ultimate goal is to develop a coherent framework that can be used in the design of optimized imaging hardware with embedded algorithms. The project is closely related to the currently popular method of super-resolution using a single image and in the context of deep learning of artificial intelligence.

Download here: Optimized Sampling Approaches for Compressive Sensing

Benjamin Rise

‘Distributed Machine Learning for Automatic Annotation and Analyses of Vast Distributed Image Archives’

This PhD project is to tackle the challenge of efficiently analysing the content of vast volumes of high-resolution imagery with distributed machine learning techniques such as federated learning. The images are generated with emerging sensors and stored in locations that span the Earth. Were it possible to bring all the imagery to one central location, it would be possible to use centralised machine learning to auto-annotate the imagery and thereby generate a list of geo-temporally localised objects of each of many types.

Download here: Distributed Machine Learning for Automatic Annotation

Will Pearson

‘Implementing Advanced Image Analytics in a Compressed Sensing and Machine Learning Framework’

This PhD will investigate generalised advanced analysis techniques to advance the capabilities (S)TEM imaging. Where applicable, supervised machine learning methods such as ANN, CNNs and Adversarial networks can be investigated for their ability to analyse images with varying degrees of noise within a microscopic setting. The aim of this project will be to leverage the support provided by the CDT to develop a generalised machine learning model capable of analysing hundreds of images.

Jianyang Xie

‘Development of Advanced AI techniques for Human Action and Behaviour Recognition’

This project will focus on the development of machine learning and high-performance computing methods for the accurate and effective recognition of human action and behaviour. Deep learning techniques will be investigated towards high recognition performance with smaller network architecture. New computing approaches will be studied to speed up the process via GPU. The models will be developed and validated by public and private datasets.

Sarah
Askevold

'Faster Uncertainty Quantification of Hydrocodes'

The aim of this project is to develop a single, integrated approach to analysing and speeding up uncertainty quantification on complex systems. The specific system used here is hydrocodes, which are high-fidelity simulations of fluid dynamics, involving computationally expensive calculations pertaining to chemistry and physics.

Download here:  Faster Uncertainty Quantification of Hydrocodes

'Bayesian Optimisation for Ever-Improving Immersive Emulation of Engagement with Fighter Jets.'

Draken Europe needs to ensure that the "red air" component of their training with the Royal Navy, RAF, and NATO is realistic and appropriate. This project will use machine learning to understand similarities between historic training exercises and propose new scenarios. Data from simulations will be augmented with synthetic data, and AI techniques like Bayesian Optimization will be used to recommend which simulations or real-world exercises to run next. The project will focus on distributed computational resources to process large datasets and develop models for articulating similarity between engagements.

Download here: Bayesian Optimisation for Ever-Improving Immersive Emulation

'Learning Transparent Models from Data-driven Algorithms to Enhance Streaming Data Analysis'

The focus of the project is to enhance algorithms for streaming data. In a world that is becoming more and more data orientated, vast amounts of data arrive in continuous streams. Prior knowledge about phenomena that give rise to this data is desirable, as this can be used to train algorithms to more effectively process future data in real time.

Download here: Learning Transparent Models from Data-driven Algorithms

'Developing Efficient Numerical Algorithms Using Fast Bayesian Random Forests.'

Decision Trees are used in data science to estimate parameter values to classify data sets in terms of regression and discrete labels. Current attempts to estimate the parameters use methods like Bayesian Additive Regression Trees (BART) and Classification Additive Regression Trees (CART), relying on Bayesian models, such as Markov Chain Monte Carlo (MCMC), which are arguably less sophisticated than what can be achieved by using Random Forests. MCMC, despite having generalised variants such as No-U-Turn-Samplers (NUTS), is hard to use in applications where the number of parameters is often unknown, making it unusable for looking at trees where the number of nodes, branches and thus parameters changes every iteration.

Download here: Developing Efficient Numerical Algorithms Using Fast Bayesian Random Forests

'Using Machine learning to train a digital test pilot for missions in turbulent environments.'

This project focuses on modelling and simulation activities for helicopter to ship operations. The objective is to develop a mathematical model of a human test pilot operating a helicopter to a ship. This digital test pilot will be trained using data gathered during real-world sea trials. It could then be used to conduct multiple virtual deck landings in order to establish the likely boundaries of a helicopter safe operational envelope. The exploitation of real-world test data investigated has the potential to enable exploration of a wider operational envelope and reduce cost and risks involved in real-world flight trials. 

Download here: Using Machine learning to train a digital test pilot

'Reinforcement Learning for Physically Aware Cyber Defence.'

This PhD project is trying to apply reinforcement learning algorithms in cyber defence scenarios while taking into consideration the impact it has on the real world. Currently, most decisions made in the context of cyber defence is to maximise availability of services on a network, as the assumption is this would increase the ability for the network to be used in such a way that it has some real-world effect.

Download here: Reinforcement Learning for Physically Aware Cyber Defence

'Algorithms and Decision-Making Processes in Distributed Attacker-Defender Games'

The project will consider a few missions from the air combat domain and map these to simpler/abstracted ‘canonical’ problems; these will form the focus for the early research. Ideally, the research will yield methods/algorithms that can be usefully mapped across to illustratively complex air-combat situations; later research will then focus on investigating these within a suitable configured game-derived simulation. 

Download here: Algorithms and Decision-Making Processes in Distributed Attacker-Defender Games

'Exploring Efficient Automated Design Choices for Robust Machine Learning Algorithms.'

This project is relates to applying Machine Learning (ML) tht currently requires the data scientist to make design choices. Different design choices can alter both how many hyper-parameters (e.g. neuron weights or kernel widths and cross-covariance terms) need to be considered but also how challenging it is to optimise the hyper-parameters of the ML algorithm. Since practitioners have limited time to perform sensitivity analyses with respect to these parameters, design choices are typically based on estimated performance with very limited consideration for the variance in this estimate. Robust performance requires that we do not optimise the hyper-parameters (e.g. using stochastic gradient descent) but generate a set of samples for the hyper-parameters that are consistent with the data and then average across these sampled values for the hyperparameters. 

Download here: Exploring Efficient Automated Design Choices

'Maximising Detection Performance Using High Performance Processing of Multi-Sensor Data.'

This project will develop state-of-the-art in signal processing algorithms to be used in submarine warfare.  Future sensing of submarines will involve robotic submarines and surface-ships as well as autonomous sensors deployed on the seabed.  To avoid the potential to give away the position of the sensors and for the ability of a submarine to avoid detection, passive listening signals using an array of hydrophones is preferable to active sensing. 

'Using machine learning and artificial intelligence to improve the tracking of vessels in sonar spectrograms.'

This PhD project explores creating an AI model that can correctly classify quiet targets in waterfall (sonar) data. Currently, waterfall data is analysed by human operators; however, this is time-consuming and expensive; these human operators outperform traditional automated passive contact follower algorithms, such as the Kalman and Alpha-Beta filters: these filters are susceptible to the abundant underwater noise and struggle with crossing tracks and quiet contacts. In contrast, humans can use their experience to learn how to mitigate the challenging aspects of the task. An automatic detection and tracking model that is more accurate and robust than traditional methods would reduce the human operator’s workload.