UKRI Centre for Doctoral Training in Artificial Intelligence, Machine Learning & Advanced Computing


Additional research projects for 2021 recruitment


ABERYSTWYTH UNIVERSITY

Project title: Prediction of facial growth for children with cleft lip and palate using 3D data mining and machine learning

1st supervisor: Dr Richard Jensen
2nd supervisor: TBD
Department/Institution: Department of Computer Science
Research theme: T2 - biological, health and clinical sciences T3 - novel mathematical, physical and computer science approaches

Project description: Approximately 150 children are born in England and Wales each year with complete unilateral cleft lip and palate (cUCLP). Despite improvements in clinical outcomes in the UK over the past 15 years, between 20-25% children with cUCLP have poor facial growth compared to 3% of the non-cleft Caucasian population. Poor facial growth results in poor aesthetic appearance and poor dental occlusion which can negatively impact on a child's psychosocial development with long-lasting effects. It is not clear why only some children with cUCLP have poor growth, nor why facial growth outcomes vary between surgeons and centres. A number of explanations have been advanced including extrinsic factors such as poor surgery in cleft palate repair during infancy, surgical technique and timing, and intrinsic factors such as the congenital absence of the upper lateral incisor, or the shape of the infants' upper arch, indicating a genetic cause. The relationship of the upper dental arch to the lower arch reflects mid-face growth and can be assessed as early as 5 years using the 5 year index. Children with cleft lip and palate in the UK have been treated in regional specialist centres since 2000 and facial growth is routinely assessed between the ages of 5 and 6 years in this way. It is also routine for cleft centres to take and keep a dental model of the upper arch of infants with cUCLP before they have any surgery. This project would involve the development of techniques for both 3D data mining and machine learning for the scanned models of infants with cUCLP, in order to determine which features are most predictive of facial growth outcome and if a predictive model can be learned. The maxillary arch models taken from infants prior to their first surgical procedure will be used along with the 5 year index score to develop models via machine learning and identify important regions. In particular, the identification of an intrinsic neonatal arch shape that is predictive of detrimental facial growth would give an opportunity to explain prognosis and manage expectations more easily with parents. It would also facilitate research on the development of new techniques for earlier treatment of poor facial growth and more personalised care for individual patients.


Project title: Few-shot Learning for Environment Adaptive Multi-modal Vision System

1st supervisor: Prof. Jungong Han
2nd supervisor: Prof. Qiang Shen
Department/Institution: Department of Computer Science, Aberystwyth University
Research theme: T3 - novel mathematical, physical and computer science approaches

Project description: As a dominating technique in AI, deep learning has been successfully used to facilitate a multitude of visual tasks, such as recognizing faces, tracking emotions or monitoring physical activities. However, each of these tasks requires training a neural network on a very large image dataset specifically collected and annotated for that task. Though the trained networks are experts for the target task, they only understand the 'world' experienced during training and can 'say' nothing about other content, nor can they be adaptive to other tasks without retraining. Moreover, most visual algorithms are learning from 'single modal', but pay no attention to other vision modalities, such as depth and thermal sensors. The core objective of the project is to develop the next generation of machine learning algorithms that can mimic human vision and intelligence, continually learning to adapt to the new environment from a few visual shots without requiring the traditional 'strong supervision' of a new dataset of each new task. Compared to the conventional supervised setting, learning from few shots poses several challenges due to, e. g. insufficient training data in a new environment; the bias between old and new tasks; the continually emerging tasks; the catastrophic forgetting problem when learning new tasks. In this project, the prospective student will solve the above problems in three stages: 1) based on our recent findings of self-paced (curriculum) learning for image classification [1] and saliency detection [2], we address the lack of annotated training data problem for real-life semantic segmentation task; 2) incorporating self-paced learning into the framework of Meta-learning, aiming to be environment adaptive by learning emerging new tasks data from old tasks; 3) delivering, distilling and reasoning semantic and geometric information over a multitude of visual data, e.g. RGB videos, depth videos and thermal videos.

[1] L. Xiang, G. Ding and J. Han, Learning from Multiple Experts: Self-paced Knowledge Distillation for Long-tailed Classification, in proceeding of European Conference on Computer Vision (ECCV 2020), spotlight paper (top 5%).
[2] D. Zhang, H. Tian and J. Han, Few-Cost Salient Object Detection with Adversarial-Paced Learning, in proceeding of 34th Conference on Neural Information Processing Systems (NeurIPS 2020).


Project title: A deep learning framework for agricultural plant breeding that predicts genotype-phenotype associations

1st supervisor: Dr Martin Swain / Dr Chuan Lu
2nd supervisor: Prof John Doonan
Department/Institution: Department of Computer Science, Aberystwyth University
Research theme: T2 - biological, health and clinical sciences T3 - novel mathematical, physical and computer science approaches

Project description: Deep learning (DL) has currently had little impact on the genetics and breeding of agricultural plants like forage crops. It is, however, desirable that the benefits of deep learning are made available to this area because half of the British landscape is dedicated to the rearing of livestock that depend on these crops. Recently, by exploiting evolutionary and functional relationships between genes, Washburn et al (2019) have developed a unique DL strategy using a convolutional neural network (CNN). After training with relevant gene expression data sets, it is able to predict expression in maize (a type of grass) under a given condition (drought) or in a particular tissue type (leaves, roots, stem etc). We propose to apply this approach to forage crops available in IBERS for which equivalent data sets are already available. First, we will address the interpretability of the model, so that we have more insight into how the CNN functions. For instance, the DeepLIFT (Deep Learning Important FeaTures) method is able to decompose the output prediction of a neural network on a specific input by back-propagating the contributions of all neurons in the network to every feature of the input. The next steps will be to improve predictive accuracy by incorporating complementary data from epigenomics or metabolomics experiments, or by incorporating prior knowledge from public databases of model plant systems. The project outputs would be directly relevant to plant breeding programmes in IBERS and beyond. This is a very data intensive project: the raw data sets involved are 100s of Gigabytes or Terabytes in size, and will require analysis using a high performance computing cluster before their use in training and testing the CNN.


Project title: Modelling seafloor change

1st supervisor: Dr Hannah Dee
2nd supervisor: Prof Jan Hiddink
Department/Institution: Department of Computer Science Aberystwyth University and Ocean Science Bangor University
Research theme: T2 - biological, health and clinical sciences T3 - novel mathematical, physical, and computer science approaches

Project description: The purchase of a new underwater robot (a Sparus II submarine from Iqua Robotics) with 2- and 3-dimensional imaging capabilities opens up new avenues for detailed spatio-temporal monitoring of sea floor state via robotic sensing. This project would be interdisciplinary, developing skills in AI, robotics, and marine science. The robot will be programmed to follow a survey path at a distance to the seafloor, and capture datasets (sidescan SONAR and visible spectrum imaging) of the same area at multiple time points. These datasets would be large and unique, with computational challenges to solve in three main areas:
- Registration, both between modalities and across times.
- Object detection and recognition in underwater domains.
- The extraction of image features from 2d and 3d which correspond to change in the marine environment.
The areas selected for survey would be busy marine environments. When a vessel is anchored the mooring chain may be dragged repeatedly across the seabed in an arc around the anchor as the vessel moves with tidal motion (Davis et al. 2016). This chain-scour is likely to impact the sediments and it is necessary to quantify the impacts of this anchor footprint on benthic communities. Broad et al. (2020) found that 90% of studies of anchoring and mooring focused on recreational vessels and identified that empirical investigations of anchor scour stemming from large merchant ships are scant and require urgent attention. The aim of this research is to gain an understanding of the effects of disturbance by anchoring activity of merchant vessels on the soft-sediments at Red Wharf Bay.
First the student will develop a method for using AIS Ship Data to estimate anchoring intensity. We will use AIS vessel tracking data (obtained through an existing collaboration with Global Fishing Watch) to identify the area impacted by anchors and anchor chains. Anchored vessels in tidal areas will move around the anchor in a half or full-circle, which means that the position of the anchor and the paths of the anchor chain can be inferred from the circle of AIS locations that is transmitted by anchored vessels. Areas with different intensities of anchoring activity will then be surveyed four times using the Sparus II AUV over the period of a year to quantify differences in sediment topography. We expect sediment topography to change throughout the year in anchored sites, and during storms in unanchored sites. Repeated surveys using side scan sonar and video Compare difference in topography between surveys Relate difference to anchoring activity between surveys If possible, capture imagery of anchoring in action, which is impossible using other methods because you cannot get close using a boat or divers. The project would enable new work to be done in underwater imaging and also marine science.


CARDIFF UNIVERSITY

Project title: The hunt for Kilonovae and multi-messenger astronomy

1st supervisor: Dr Cosimo Inserra
2nd supervisor: TBD
Department/Institution: School of Physics and Astronomy, Cardiff University
Research theme: T1 - data from large science facilities

Project description: Core-collapse supernovae are the final, explosive demise of massive stars and are responsible for black hole and neuron star formation. General Relativity tells us that their binary orbit will shrink owing to energy losses via gravitational-wave (GW) emission. Following this shrinking, during the final few orbits, a prominent gravitational waveform is produced. Mergers of compact binaries, therefore, represent the true final fates of massive stars and are the dominant source of all heavy elements in the Universe.
Kilonovae are the electromagnetic counterparts created by collisions between neutron stars, the dense products of collapsed stars. These are currently the only sources responsible for gravitational waves and electromagnetic radiation, making them a unique 'multi-messenger' probe of the Universe. Understanding these sources will reveal whether neutron star mergers produce the full cosmic budget of chemical elements heavier than iron, and how matter behaves at the most extreme densities possible. At the moment only one kilonova, associated with GW170817, has been observed but rate estimates predict a few events in the next LIGO runs up to a few dozen in LIGO A+ (starting in 2024). This is sufficient to provide constraints on the nature of nuclear matter.
The project will focus on the electromagnetic follow-up of neutron stars mergers (and if discovered NS-BH mergers counterparts) and it will be carried out as part of the ENGRAVE collaboration, the LIGO/VIRGO Collaboration and the upcoming Legacy Survey of Space and Time (LSST) at the Vera Rubin Observatory. One of the goals of the project will be building a multi-messenger code to fit both gravitational wave and electromagnetic data simultaneously which is pivotal to deliver the first multi-messenger statistical study of neutron star mergers. Machine learning approaches will also be used to retrieve any link between the electromagnetic information and those retrieved from the gravitational waves analysis and modelling. If the dataset will be rich enough, an Artificial intelligence algorithm can be built to identify kilonovae candidates during the downtimes between LIGO runs.
In this project, the PhD student will gather knowledge of GW physics, time-domain astronomy, astrophysics of exploding/merging cosmic objects and stellar evolution. From the computational side, the student will acquire programming skills in python and machine learning / Artificial Intelligence techniques and environments. Experience in observational astronomy and statistics are ‘de facto’ outcomes of such project.


Project title: Exploring the environment of black holes merger and its connection with gravitational waves

1st supervisor: Dr Cosimo Inserra
2nd supervisor: TBD
Department/Institution: School of Physics and Astronomy / Cardiff University
Research theme: T1 - data from large science facilities

Project description: Core-collapse supernovae are the final, explosive demise of massive stars and are responsible for black hole formation. As a consequence of the prevalence of binarity amongst massive stars, they provide the leading progenitor channel of producing compact object binary systems with two black holes. General Relativity tells us that their binary orbit will shrink owing to energy losses via gravitational-wave (GW) emission. Following this shrinking, during the final few orbits, a prominent gravitational waveform is produced. Mergers of compact binaries therefore represent the true final fates of massive stars. However, unlikely mergers where a neutron star is involved, there is no electromagnetic emission arising from the merger of two black holes due to their intrinsic nature. Hence, any effort to link any gravitational waveform produced by a black holes merger to astrophysical information and/or its progenitor stars has produced null results.
The LIGO/Virgo collaboration has recently entered its third observing season and up to nowadays, they detected more than 20 gravitational wave events from merging binary black hole systems. More will come in the future thanks to the improved sensibility. Cardiff University is full member of the LIGO/Virgo collaboration and has preferential access to the gravitational-wave discovery and their information. Cardiff University is also part of the largest European endeavour in gathering astrophysical information from any compact merger producing GWs (ENGRAVE collaboration). Hence, the PhD student taking up this project will be in a unique position to get the most from two different approaches and collaborations.
The project will focus on the environment of black-hole mergers, which is the only way to retrieve useful astrophysical information from such events. The project will focus on retrieving information on galaxies in the likelihood region of previous GW mergers via electromagnetic spectroscopy (for the distant events) and integral-field spectroscopy (for the closest events). The latter allows for a spatially-resolved investigation of the surrounding stellar populations and provide constraints on the formation scenario of the binary and the 3D position of the merger in the galaxy harbouring the gravitational-wave event. Machine learning approaches will then be used to retrieve any link between the environmental information and those retrieved from the waveform. If the dataset will be rich enough, an Artificial intelligence algorithm can be built to predict what kind of galaxy will likely be the host of future, far black hole mergers.
In this project, the PhD student will gather knowledge of GW physics, astrophysics of galaxies and stellar evolution. From the computational side, the student will acquire programming skills in python and machine learning techniques and environments. Experience in observational astronomy and statistics are 'de facto' outcomes of such project.


Project title: Plus ultra: extreme supernovae beyond the standard paradigm of cosmic explosions

1st supervisor: Dr Cosimo Inserra
2nd supervisor: TBD
Department/Institution: School of Physics and Astronomy / Cardiff University
Research theme: T1 - data from large science facilities

Project description: Supernovae (SNe), stellar explosions staging the final act of a stars life, play an important role in many astrophysical domains, for instance stellar evolution, feedback in galaxy formation, synthesis and distribution of almost all the elements and raw materials for both star and planet formation.
The last ten years, with the advent of wide-field surveys, have opened up a new parameter space in time-domain astronomy with the surprising discovery of transients defying our understanding of how stars explode. These can be grouped into three categories: 1 a population of ultra-bright superluminous supernovae, some 100 times brighter than classical supernova types, offering new probes of the high redshift universe and the potential for a new class of standard candle; 2 transients showing fast rise and subsequent rapid decay that do not resemble any common class of extragalactic transient; 3 transients with extreme energetics or complex evolution happening in low-metallicity or low-luminosity environments.
The consensus is that metallicity, initial mass and multiplicity influence the type of SN we observe but their precise role has not been characterised. This impedes our ability to use SN as probes of star formation across the Universe and to understand if the progenitor metallicity is encoded within the mass and explosion mechanism of such extreme supernovae. Detailed analyses of such SNe and their hosts, both in their local (supernova position) and global (host galaxy as a whole) environment, have generally been restricted to a few, usually luminous, host galaxies. Machine learning approaches will then be used to retrieve any link between the environmental information and those retrieved from the supernova. If the dataset will be rich enough, an Artificial intelligence algorithm can be built to predict what kind of galaxy will likely be the host of future extreme Supernovae and to recognize them early enough in the 22 Terabyte stream of data (per observing night) that the Vera Rubin observatory will deliver from 2023.
In this project, the PhD student will gather knowledge of supernova explosions linked to the environment properties as well as programming skills in python, machine learning / Artificial Intelligence, experience in observational astronomy and statistics.


Project title: Exploring the Gravitational-Wave Sky with Machine Learning

1st supervisor: Patrick Sutton
2nd supervisor: Stephen Fairhurst
Department/Institution: Astrophysics PHYSX
Research theme: T1 - data from large science facilities

Project description: Gravitational waves (GWs) are produced by some of the most violent events in the Universe, such as the mergers of black holes and the explosive deaths of massive stars. Rapid detection of such signals can trigger follow-up observations by other facilities including ground-based telescopes, satellites, and neutrino observatories, greatly increasing the scientific payoff of such discoveries. For example, combined observations of the binary neutron star merger GW170817 revealed the origin of heavy elements in the Universe and provided a new way to measure the Hubble constant.
Detecting such GW events before the electromagnetic counterpart fades requires analysis of the gravitational-wave data on timescales of minutes or less. Recently, machine learning techniques based on Convolutional Neural Networks (CNNs) have been demonstrated to have the potential for sub-second-latency analysis of data from GW detector networks. The aim of this project is to develop and implement a CNN with the ability to detect generic transient signals, such as those expected from newly formed or perturbed black holes and neutron stars. The analysis will be run in real time during upcoming observing runs by the LIGO-Virgo-KAGRA detector network, to detect, characterise, and determine the location on the sky of GW signals. We will issue public alerts about detected events, with special emphasis on signals associated with gamma-ray bursts, core-collapse supernovae, and other relativistic astrophysical phenomena. This project involves developing expertise in programming, signal processing, high-throughput computing, and high-performance computing. You will collaborate with other GW observers, GW theorists, and astronomers in the Cardiff University Gravity Exploration Institute and in the LIGO Scientific Collaboration. The project may optionally involve a long-term (up to four months) secondment to one of the LIGO observatories in the US.


Project title: Simulation-based Inference of gravitational waves signals from black holes and neutron stars.

1st supervisor: Vivien Raymond
2nd supervisor: Stephen Fairhurst
Department/Institution: School of Physics and Astronomy / Cardiff University
Research theme: Gravitational-wave Astrophysics

Project description: Black holes and neutron stars are the densest objects in the universe, well beyond what we can produce in a laboratory and at the very edge of our understanding of physics. They lead to puzzling physical consequences, in particular regarding the behaviour of space and time. When they collide, they produce the most violent events in the universe, shaking space and time and creating gravitational waves: ripples in the fabric of spacetime which propagate away at the speed of light. Gravitational-waves were observed for the very first time in September 2015, when two colliding black holes were detected by the LIGO-Virgo collaboration. Since then, several signals have been observed, and we were able to characterise the black holes and neutron stars at the source of those gravitational waves.
This characterisation currently involves stochastic sampling methods with a very high computational cost, and simplified assumptions of the detectors properties. This project will leverage modern advances in likelihood-free inference methods, and in particular simulation-based inference, to solve this inference problem accurately. Our approach will adapt automatically to changing features in the detector noise, allow for new data to be continuously included, and will be applicable to the upcoming new generation of gravitational-wave detectors.
Gravitational-wave sources are laboratories where we can measure in neutron stars the equation of state of matter at densities otherwise unattainable, and test General Relativity in the strong field regime. Inference of their extragalactic population enables new understandings of the Universe's structure of matter, and independent measurements of the Universe's expansion.


Project title: And yet it moves... developing innovative physics-aware machine learning techniques to teach computers how galaxies spin

1st supervisor: Dr Timothy A. Davis
2nd supervisor: Dr Edward Gomez
Department/Institution: School of Physics and Astronomy, Cardiff University
Research theme: T1 - data from large science facilities (Astrophysics)

Project description: Understanding the formation and evolution of galaxies is a vital part of modern astrophysics. A key challenge, however, is that the physics that matters happens over a vast range of scales. From central black holes to vast dark matter halos, only by obtaining a robust picture across galaxies can we truly understand the physical processes driving galaxy evolution. The kinematics of the gaseous components of galaxies provide a key probe of these physical process. In the upcoming decade, next-generation observatories will reveal such motions in millions of galaxies, but our tools for interpreting this data are not fit for the "big data" era. A student taking on this project would develop innovative fast, unsupervised (or self-supervised) kinematic modelling techniques based on physics-aware convolutional auto-encoders, and apply them to new data from state-of-the-art telescopes around the world to help usher in the new era of data-intensive astronomy.


Project title: Investigating scenario generality through AI for self-parking vehicles

1st supervisor: Prof Roger Whitaker, Cardiff University
2nd supervisor: Dr Liam Turner, Cardiff University
Department/Institution: Cardiff University School of Computer Science and Informatics
Research theme: T3 - novel mathematical, physical and computer science approaches

Project description: It has now been established that evolutionary approaches to constructing neural networks are effective in performing complex simulated self-driving tasks, such as parking. This represents a form of reinforcement learning: for a nice demonstration example see here.
In this project we use an extension of the vehicle parking problem to consider effective strategies for scenario generalisation - for example, being able to park a more complex vehicle (e.g., an articulated vehicle with a trailer) in any given location, without necessarily being highly trained on every specific possible scenario in advance.
Using extensions to the neuro-evolutionary algorithms such as NEAT, we will consider how transferable learning can be established and applied through the parking scenario. We will use extensive 2D simulation, with the possibility of representing the work in tools such as Unity for demonstration purposes. This PhD is suitable for someone keen to gain in-depth knowledge of state-of-the-art deep reinforcement learning through evolutionary processes. These have been established as important techniques in the context of self-driving.


Project title: Topological evolution of Neural Networks through network building blocks.

1st supervisor: Dr Liam Turner, Cardiff University
2nd supervisor: Prof Roger Whitaker, Cardiff University
Department/Institution: Cardiff University School of Computer Science and Informatics
Research theme: T3 - novel mathematical, physical and computer science approaches

Project description: Research surrounding the topological evolution of artificial Neural Networks (NNs) have found particular efficacy in promoting modularity and regularity as the networks evolve [1,2], mimicking processes found in nature. This project will aim to go a further in this direction by examining how the composition of particular substructures and sequence signatures of these [3], seen as the building blocks of complex networks in nature [4], manifest in the evolution of NNs and determine how targeting preservation or disruption of this effects learning.

[1] Stanley, K. O., D'Ambrosio, D. B., & Gauci, J. (2009). A hypercube-based encoding for evolving large-scale neural networks. Artificial life, 15(2), 185-212.
[2] https://www.youtube.com/watch?v=FUqYNRZTl3U
[3] Milo, R., Shen-Orr, S., Itzkovitz, S., Kashtan, N., Chklovskii, D., & Alon, U.((2002). Network motifs: simple building blocks of complex networks. Science, 298(5594), 824-827.
[4] Milo, R., Itzkovitz, S., Kashtan, N., Levitt, R., Shen-Orr, S., Ayzenshtat, I., ... & Alon, U. (2004). Superfamilies of evolved and designed networks. Science, 303(5663), 1538-1542.


Project title: Epigenetic evolution of Neural Networks.

1st supervisor: Dr Liam Turner, Cardiff University
2nd supervisor: Prof Roger Whitaker, Cardiff University
Department/Institution: Cardiff University School of Computer Science and Informatics
Research theme: T3 - novel mathematical, physical and computer science approaches

Project description: In evolutionary learning of neural networks, survival of the fittest is an overarching mechanism to drive progressive performance. This results in the majority of models being discarded in pursuit of optimising against specific objectives. When a trained model is then used as the basis for unseen problems, a new neuroevolutionary process starts without a means of drawing on the efficacy of previously discarded solutions that may be useful. This project will examine whether the characteristics of epigenetic processes [1] found in nature (e.g., biomarkers and gene expression) can inspire methods for preserving phenotype to objective history and how drawing on this can affect learning of new tasks as well as revisiting tasks.

[0] Epigenetics video
[1] Stanley, K. O., D'Ambrosio, D. B., & Gauci, J. (2009). A hypercube-based encoding for evolving large-scale neural networks. Artificial life, 15(2), 185-212.


Project title: Real-time Situational Understanding using Deep Neural Networks and Knowledge Graphs

1st supervisor: Prof Alun Preece, Cardiff University
2nd supervisor: Dr Jose Camacho Collados
Department/Institution: Cardiff University School of Computer Science and Informatics
Research theme: T3 - novel mathematical, physical and computer science approaches

Project description: DeThe problem of situational understanding involves recognising existing situations ("what's currently happening?") and drawing inferences about possible future situations ("what might happen next?"). Deep neural networks (DNNs) for activity recognition have emerged as effective technologies in recent years for recognising existing situations. Approaches such as 3D convolutional neural networks (3D CNNs) and long short-term memory networks (LSTMs) can deal with the problem of recognising situations that are changing in real-time; however, they require large amounts of training data, making them problematic for recognising rare or unusual situations. For the second aspect -- drawing inferences about possible future situations -- DNNs also suffer serious limitations, due to their restricted ability to deal with cause-and-effect and open-world (unseen) situations. Knowledge graphs offer a way to deal with these limitations, to augment DNN capabilities with background or prior knowledge.
Currently, however, effective integration of knowledge graphs with DNNs is limited and often shallow. This PhD will look at deeper ways to integrate the benefits of DNNs and knowledge graphs, applied to the problem of situational understanding. The project will be conducted in collaboration with Cardiff University's Crime and Security Research Institute, which will provide access to real-world case studies such as managing the flow of misinformation in online networks, and rapid decision-making in "front line" contexts.


Project title: Using AI to find the first galaxies

1st supervisor: Prof. Stephen Eales
2nd supervisor: Dr Matt Smith
Department/Institution: Cardiff University/ School of Physics and Astronomy
Research theme: T1 - Astronomy/AI

Project description: The origin of galaxies is one of the biggest questions in astronomy. The biggest step towards answering this question came with the first submillimetre surveys in the late nineties, which revealed a population of luminous dusty galaxies ten billion years in the past which are almost completely hidden by the dust from traditional optical telescopes, even one as powerful as Hubble. These galaxies are forming stars so quickly that we are effectively seeing galaxies formed in front of our eyes. We have so far been able to find galaxies like this, using submillimetre telescopes like the Herschel Space Observatory, out to a redshift of about six, roughly a billion years after the big bang.
We now have an opportunity to go even further back in time. Our group has built a new millimetre camera that will be installed on the Large Millimetre Telescope in Mexico later this year. Surveys with this camera will allow us to find galaxies within a few hundred million years of the big bang.
The big challenge, however, with all these surveys is that the poor angular resolution makes is very difficult to identify the galaxies producing the submillimetre radiation. Our big Herschel survey, for example, discovered half-a-million submillimetre sources, but we have only been able to identify the galaxies responsible for the submillimetre radiation for about half of these. The answer may be AI. By using training sets obtained from small areas of sky where the observations are exceptionally good and we have been able to identify all of the submm-emitting galaxies, we plan to train neural networks and deep learning techniques to identify the submm-emitting galaxies where the observational data is not so good. The student will work on these techniques but also on the many observational programmes, using telescopes on the ground and in space, that we have underway to study these objects – galaxies in the process of formation.


SWANSEA UNIVERSITY

Project title: Predicting Effective Control Parameters for Evolutionary Algorithms Using Machine Learning Techniques

1st supervisor: Dr Sean Walton
2nd supervisor: Dr Alma Rahat
Department/Institution: Computational Foundry, Swansea University
Research theme: T3 - novel mathematical, physical and computer science approaches

Project description: Evolutionary algorithms have proven to be capable of solving complex problems. A limitation of these techniques is that they often have a number of control parameters which need tweaking for every new problem they are applied to in order to maximise performance. The most common way to address this problem is to allow these control parameters to adapt as the algorithm runs. Recent developments however have shown that by sampling a new problem it is possible to predict effective control parameters before the start of the optimisation. In this project you will work to develop new techniques using machine learning to predict effective control parameters for new problems.


Project title: Mixed-initiative procedural content generation for level design in video games

1st supervisor: Dr Sean Walton
2nd supervisor: Dr Alma Rahat
Department/Institution: Computational Foundry, Swansea University
Research theme: T3 - novel mathematical, physical and computer science approaches

Project description: Game developers are under increasing pressure not only to launch games with hours of unique content, but to continue to add new fresh content post launch. Creating new and diverse content is expensive both in terms of time and money which provides motivation to develop tools to support them. Mixed-initiative procedural content generation algorithms seek to support game and level designers in the creative process. In this project you will work to develop new techniques for supporting game designers design levels and other content for video games.


Project title: Adverse outcomes in infancy and childhood: explicating multifactorial aetiologies through AI

1st supervisors: Prof Sue Jordan (Nursing), Dr. Ben Mora (Computer Science)
Supervisory team: Dr Adam Turner (pharmacy), Neil Carter (informatics), Darren Edwards (statistics)
Department/Institution: Health and Life Science (Nursing, Pharmacy, Medicine/ Computer Science, Swansea University
Research theme: T2 - biological, health and clinical sciences

Project description: Adverse perinatal and childhood outcomes and use of prescription medicines are concentrated in the most deprived communities. However, aetiology is complex and multifactorial. We aim to use machine learning and AI techniques to identify predictors, starting in early pregnancy, of neurodevelopmental delay, and serious ill-health. This may help support in medicines management, childcare and breastfeeding to be targeted most effectively. Machine learning techniques applied to the linked data in SAIL will allow us to identify dyads 'at risk' based on patterns and combinations derived from comprehensive characterisation of the range of variables affecting childhood outcomes. Data available include: in utero and breastmilk exposures to medicines (with prescription dates) and disease, smoking, alcohol intake, substance misuse (with duration), healthcare contacts, vaccination status, socio-economic status (SES), rurality, distance from landfill and incinerators, and infections in pregnancy (including COVID-19 in 2020). This project will run alongside the Innovative Medicines Initiative Conception project (2019-24). There are opportunities to participate in this larger group, and share findings. To validate and contextualise the findings of the machine learning work, we shall undertake semi-structured interviews with stakeholders and healthcare professionals.


Project title: Minimising optimising prescription regimens from known ADRs: pattern recognition using AI

1st supervisors: Dr. Ben Mora (Computer Science), Prof Sue Jordan (Nursing)
Supervisory team: Dr Adam Turner (pharmacy), Neil Carter (informatics), Darren Edwards (statistics)
Department/Institution: Health and Life Science (Nursing, Pharmacy, Medicine/ Computer Science, Swansea University
Research theme: T2 - biological, health and clinical sciences

Project description: The WHO's third global safety challenge is reduction of medicines-related harm. However, the most harmful combinations of prescription medicines, medical diagnoses, signs and symptoms of illness and/or adverse effects of medicines, and contextual variables remain largely unknown. Supported by AWTTC (All Wales Therapeutics and Toxicology Centre) and the Royal Pharmaceutical Society, we have developed a comprehensive list of signs and symptoms putatively associated with primary care medicines. Several neural networks applied to the linked data in SAIL will be investigated to identify patterns of these signs and symptoms (as recorded), prescriptions, and diseases most closely associated with adverse outcomes. Currently, juxtaposing patients' signs and symptoms with full lists of prescribed medicines challenges healthcare professionals, and application of AI and machine learning will offer clinically useful 'short cuts' to harm recognition in practice. This project will run alongside ongoing impact casework and a funded clinical centenary scholarship. To validate and contextualise the findings of the machine learning work, we shall undertake semi-structured interviews with stakeholders and healthcare professionals.


Project title: Putting the Human Back in the Loop of Bayesian Optimisation for Aerospace Design

1st supervisor: Dr Alma Rahat
2nd supervisor: Dr Sean Walton
Department/Institution: Computational Foundry, Swansea University
Research theme: T3 - novel mathematical, physical and computer science approaches

Project description: Bayesian optimisation algorithms have proved useful for solving many complex aerospace design problems. In simple terms a designer defines a problem with a starting shape for a component such as a wing as well as some kind of performance metric the algorithm tries to maximise. Computational fluid dynamic (CFD) simulations are run to estimate the performance metric to drive the optimisation process which may be computationally expensive and time-consuming. A surrogate model for performance is thus used to identify the most promising solutions that may be subjected to CFD simulations. As new simulations are performed, the surrogate model is retrained and predictions for good solutions improve, and after a certain number of simulations, the best estimation of the optimal design is presented to the designer. Most of the current research has been focused on reducing the number of simulations required to estimate the best design by introducing new utility functions that define how to balance between exploration and exploitation from the predictions of surrogate models in locating promising solutions. However, they mostly ignore the preferences from the designer on some of the most important aspects of design, such as aesthetics or fabrication feasibility, that can not be simply modelled by a computer during the optimisation process. Therefore, the successful PhD candidate will develop techniques for putting the human back into the loop and allowing the optimiser to be driven by designer preference as well as performance metrics.


Project title: Predicting Effective Control Parameters for Evolutionary Algorithms Using Machine Learning Techniques

1st supervisor: Dr Sean Walton
2nd supervisor: Dr Alma Rahat
Department/Institution: Computational Foundry, Swansea University
Research theme: T3 - novel mathematical, physical and computer science approaches

Project description: Evolutionary algorithms have proven to be capable of solving complex problems. A limitation of these techniques is that they often have a number of control parameters which need tweaking for every new problem they are applied to in order to maximise performance. The most common way to address this problem is to allow these control parameters to adapt as the algorithm runs. Recent developments however have shown that by sampling a new problem it is possible to predict effective control parameters before the start of the optimisation. In this project you will work to develop new techniques using machine learning to predict effective control parameters for new problems.


Project title: Efficient Learning of the Optimal Probability Distribution over the Policy Space in Reinforcement Learning.

1st supervisor: Dr Alma Rahat
2nd supervisor: Dr Sean Walton
Department/Institution: Computational Foundry, Swansea University
Research theme: T3 – novel mathematical, physical and computer science approaches

Project description: Reinforcement Learning is an important technique in Machine Learning for control problems, and it is closely related to how we learn in an unknown environment. In this method, an agent performs a possible action based on what state it is in and its prior belief of the reward for that action, and receives a reward or punishment as a consequence of that action; this helps it to differentiate between good and bad actions given a state as the agent may update its belief with more experience. Essentially, the learner aims to develop an estimation of the optimal probability distribution for selecting an action over the state-action space (also known as the policy space). Traditional approaches use repeated trials and errors to estimate this distribution over the policy space, which can be time-consuming due to the sheer number of required repetitions for a good estimation. Given the utility of reinforcement learning, it would be game changing to be able to improve the speed of learning by reducing the number of repetitions required. In this project, inspired from the Bayesian optimisation approaches, we propose to investigate a novel data-driven direct policy search approach where you will model the probability distribution from carefully selected data, take an entropy based search approach to identify the most informative trials to perform, and sequentially improve the estimation of the optimal probability distribution over the policy space.


Project title: Robust Parameter Optimisation for Image Segmentation

1st supervisor: Dr Thomas Torsney-Weir
2nd supervisor: Dr Alma Rahat
Department/Institution: Computational Foundry, Swansea University.
Research theme: T2: biological, health and clinical sciences (medical imaging, electronic health records, bioinformatics) and T3: novel mathematical, physical, and computer science approaches (data, hardware, software, algorithm

Project description: Medical professionals rely on robust image segmentation algorithms to highlight anomalous tissues in a patient, e.g. cancer. However, image segmentation algorithms are typically calibrated on a limited number of training images through a tedious trial and error process. This gives limited context to the robustness against overfitting which makes it difficult to predict how well the segmentation algorithm will perform on unseen images. This could lead to an incorrect diagnosis. The goal of this project is to use a combination of visualization techniques (e.g. Tuner[1]) and Bayesian model calibration techniques (e.g. history matching[2]) to develop a system for robust parameter optimisation with a focus on image segmentation algorithms.

[1] Torsney-Weir, T., Saad, A., Moller, T., Hege, H.C., Weber, B., Verbavatz, J.M. and Bergner, S., 2011. Tuner: Principled parameter finding for image segmentation algorithms using visual response surface exploration. IEEE Transactions on Visualization and Computer Graphics, 17(12), pp.1892-1901.
[2] Andrianakis, I., Vernon, I.R., McCreesh, N., McKinley, T.J., Oakley, J.E., Nsubuga, R.N., Goldstein, M. and White, R.G., 2015. Bayesian history matching of complex infectious disease models using emulation: a tutorial and a case study on HIV in Uganda. PLoS computational biology, 11(1), p.e1003968.


Project title: Multi-Objective Evolutionary Approach Towards Dynamic Graph Drawing

1st supervisor: Dr Alma Rahat
2nd supervisor: Dr Daniel Archambault
Department/Institution: Computational Foundry, Swansea University.
Research theme: T3: novel mathematical, physical, and computer science approaches (data, hardware, software, algorithm

Project description: To draw an appealing dynamic graph, it is important to strike a balance between readability and stability as there is a natural conflict between these objectives [1]. Current approaches set up force systems which can be optimised to reach a reasonable trade off between the two. In this project, we investigate how multi-objective evolutionary algorithms can be used to explore the trade-offs between readability and stability.

[1] D. Archambault and H. C. Purchase. Can animation support the visualization of dynamic graphs? Information Sciences, 330:495–509, 2016.


Project title: Visual Analytics for Public Health Network Analysis

1st supervisor: Daniel Archambault
2nd supervisor: TBC
Department/Institution: Computer Science Swansea University
Research theme: T2 - biological, health and clinical sciences

Project description: In public health settings, social networks are often encoded as multivariate graphs with both static and dynamic information. The human actors in these networks have demographic and survey information associated with them along with their social ties. Given this information, the analyst wants to understand how the information associated with the nodes and the social ties influence behaviours (under-age drinking, mental health, non-suicidal self injury etc). In this project, how network analytics and ML can be supported by visual analysis in this setting.


Project title: How to Identify Effective Policies for Managing a Pandemic?

1st supervisor: Dr Alma Rahat
2nd Supervisors: Dr Thomas Torsney-Weir, Dr Daniel Archambault
Department/Institution: Computational Foundry, Swansea University.
Research theme: T3: novel mathematical, physical, and computer science approaches (data, hardware, software, algorithm

Project description: During a pandemic, policymakers effectively use sophisticated numerical simulators, such as the model developed by researchers at the London School of Hygiene and Tropical Medicine (LSHTM) for Covid 19, to evaluate the impact of their decisions, for example imposing lockdown, on the progression of the disease at hand. Identifying which decisions should be evaluated with the computationally expensive simulator still largely depends on human intuition and domain knowledge. Considering the large decision space, in the current approach, we cannot know if there are alternative policy actions that may yield even better outcomes in terms of disease progression and economic impact of actions. In this project, building upon our current work with the Welsh Government, we will construct data-driven Bayesian surrogate models that capture the relationship between policy actions, disease progression and economic effects. Using these surrogates, we aim to locate a range of promising policy actions that optimally trade-off between disease progression and economic impact. We expect this to allow policymakers to make more informed decisions rapidly. If successful, this work will help policymakers in future public health emergencies, and other multi-objective data-driven policy problems.