Full details of the application process are available online. On your application form, please state that you wish to be considered for admission to the Centre for Machine Learning, Department of Computer Science.
Reasoning about what has been learned
Artur Garcez and Eduardo Alonso
Deep networks have become very successful recently as an efficient machine learning method for pattern recognition and classification of large-scale multimodal data (images, audio, sensor data), having obtained state-of-the-art performance in tasks such as speech recognition. The Research Centre for Machine Learning is world-leading in knowledge extraction from neural networks. Knowledge extraction enables explanation, opening up the "black-box", and reasoning towards a better understanding of the fundamental processes behind the success of neural networks. In this project, novel knowledge extraction methods and algorithm for deep networks will be investigated, developed and evaluated systematically. Knowledge will be extracted from deep (recurrent) networks trained to solve real-world big data tasks.
Transfer learning for human-like computing
Artur Garcez and Chris Child
One of the main criticisms of current machine learning is the inability to re-use knowledge from a learned task to facilitate learning of a new related task. Humans, on the other hand, are excellent at learning from a few examples only, and adapting to new situations through the use of analogy. The area of transfer learning seeks to address the problem by studying the mechanisms of knowledge and data re-use between a source and target domain. In this project novel transfer learning methods and algorithm based on knowledge extraction, analogy making and human reasoning will be investigated, developed and evaluated. Transfer learning experiments will be carried out using benchmark and real-world source and target domains.
Deep learning for cervical spine image analysis
Greg Slabaugh and Artur Garcez
The cervical spine (neck) is a highly flexible part of the spine and is particularly vulnerable to trauma. Dislocation or fracture of the cervical vertebrae have potential for long-term and life-changing disabilities. In this project, you will develop novel, automated detection and segmentation algorithms based on deep learning methods like convolutional neural networks to identify and precisely delineate the cervical vertebrae in a computed tomography (CT) image. You will additionally research shape analysis algorithms to identify any fractures, and perform further processing to characterise the alignment of the bones in the vertebral column. Experiments will be performed with 3D CT images collected from hospitals in the UK and abroad.
Fisher-Tippett Tissue Characterisation and Registration of Ultrasound Images
Greg Slabaugh and Eduardo Alonso
Ultrasound is a common medical imaging modality as it is low cost, portable, and safe. However, it is corrupted by speckle, a noise-like pattern resulting from coherent accumulation of random scattering in a cell of ultrasound beam. As a result of the speckle, the intensity of pixels in a homogenous region of an ultrasound image have been shown to follow a Fisher-Tippett distribution. In this project, you will adopt this mathematical model for characterising tissues in an ultrasound image, for example, using clustering techniques based on information theoretic distances based on Fisher-Tippett distributions. You will research and develop novel registration algorithms using these tissue models to spatially align ultrasound images using affine and non-rigid techniques. Potential applications include cardiac and foetal image analysis.
Deep Learning as a Computational Model of Classical Conditioning
Eduardo Alonso and Chris Child
Classical conditioning (CC) is a fundamental learning paradigm by which organisms adapt to their environments by associating stimuli. Although behavioural neuroscientists have proposed theories of increasing complexity that can account for a good number of CC phenomena, we are still lacking a computational framework that would characterise CC structures in a formal way, and to accurately test their predictions in simulations. In collaboration with the Centre for Computational and Animal Learning Research, the aim of this project is to use Deep Learning architectures to formulate such framework using a combined approach: first, to learn to uniquely identify stimuli and how they generalize (convolutional neural networks); then, over this representation, to learn complex associative structures (recurrent neural networks that embed reinforcement learning). We are also interested in investigating how the resulting model could inform research in learning impairments associated with neurodevelopmental disorders.
Deep Learning for Cybersecurity
Eduardo Alonso and Artur Garcez
Machine learning has been mostly applied to the identification of IDS anomalies using supervised classification techniques. In this project we will explore how unsupervised optimization methods can be applied to automate computer security audits. Such process is currently carried out by human experts, by compromising the system's security using a database of exploits - which is time consuming and prone to errors. Also, new exploits and zero-day vulnerabilities can go undetected, and multiple-attacks are typically not ignored. We plan to use Deep Learning methods to simulate these types of attacks and to learn the best possible sequence of actions to defend the system under consideration.
Model Acquisition and Planning using Deep Q-Networks
Chris Child and Eduardo Alonso
Research by Google Deep Mind in the field of Reinforcement Learning using Deep Learning Neural Networks has received intense media attention in recent months. Deep Q-Networks, based around Deep Learning and Reinforcement Learning, have been demonstrated to be able to play Go at the level of a master and to play Atari 2600 games without human intervention. The aim of this research is to develop an extension to Deep Q-Networks by acquiring a model of an environment using features extracted by the network. At each moment in time, the neural network produces a set of features from the environment inputs in its later layers. The system should model these features, predicting which ones will be present in the next time step. The addition of model acquisition to Deep Q-Networks is expected to contribute to the technique by: (i) increasing the speed of learning (increasing simulation speed using the model rather than the real environment); and (ii) improving behaviour in previously unseen environment states by using the predictive capabilities of the model (planning).
Applying Reinforcement Learning to Behaviour Trees for Computer Game Agents
Chris Child and Artur Garcez
Behaviour trees are a flexible, modular and scalable architecture to handle AI in computer games, providing a top-down approach to split high level behaviours into smaller tasks in an intuitive manner. They have become popular as a replacement for finite state machines (FSM), which are ubiquitous in the games industry. This research will enhance the decision making process of behaviour trees by applying approximate dynamic programming or reinforcement learning to improve performance. Due to the real time nature of virtual environments, algorithms will be developed that use offline learning, combined with online-learning run in parallel, or over multiple time-slices. This research will provide methods for implementing agent behaviour that can be predictably restricted within defined parameters, whilst maintaining the level of flexibility and scalability that behaviour trees provide. The developed technology could be applied to a number of different game domains, considerably reducing human effort involved in the design process.
Static Analysis for Symbolic Machine Learning
Jacob Howe and Eduardo Alonso
Static program analysis often involves calculating a fixed point in a lattice: as information about values at a program point is determined an approximation to the possible range of values is refined and propagated. Similarly, a variety of machine learning techniques work by working through a training dataset until a classifier is learnt which defines a frontier in such a lattice. This project aims to explore and exploit the synergy between fixed point techniques developed for static analysis and symbolic machine learning techniques such as Inductive Logic Programming towards novel machine learning algorithms and systems.
Variable Elimination and Static Program Analysis with Application to Knowledge Extraction
Jacob Howe and Artur Garcez
The development of secure and safe software is strengthened by a final step of verification, in the sense of formally establishing that a precisely specified class of bugs does not occur. Abstract interpretation provides a formal framework for static program analysis using operations on an abstract domain. The aim of this project is to exploit recent work on variable elimination in the abstract domain of polyhedra to provide a library of domain operation and apply it in a range of program analysis contexts. The techniques and tools developed in this context can be applied to the problem of knowledge extraction from neural networks, thus contributing to "opening the black box" and addressing the lack of interpretability of neural networks, a common criticism of such learning systems.
Multi-timescale sequence models for text and music with recurrent neural networks
Tillman Weyde and Eduardo Alonso
Sequence models for music and language have recently received a boost from advances in recurrent neural network. Although in principle unbounded, recurrent neural networks like most sequence models, do well with local structure, but capture less of long term dependencies. Different architectures like LSTM, SCRN and hierarchical models have been proposed, but there is significant room for improvement. Improved models will be developed and evaluated on speech and music tasks, such as text prediction and musical rhythm and melody classification and generation.
Multidimensional Sequence Models for Audio and Music with Hybrid Neural Oscillator Networks
Tillman Weyde and Artur Garcez
Modelling high-dimensional phenomena, such as audio and music, is a challenge for learning models, but has immediate benefits for audio and music analysis. Applying hybrid models including oscillators and recurrent network layers can help but modelling temporal regularities. Music transcription and general event detection tasks will be the applications of improved models.