Bridging the Gap

Bridging the Gap: Fellowships

Fellowships

Current Calls

Calls for BTG fellowships are now closed.

Second Fellowship Award Winners

The applications to the second round of fellowships were of such high quality that we decided to award a three fellowships in this round.

A Framework for Dynamic Self-Optimization of Power and QoS in Cloud Architectures Rami Bahsoon, School of Computer Science

Cloud computing is claimed for enabling convenient, on-demand network access to a shared pool of configurable computing resources (e.g., networks, servers, storage, applications, and services) that can be rapidly provisioned and released with minimal management effort. Various vendors have successfully provided cloud computing, among them, Amazon Web Services, Microsoft Azure, Google AppEngine, Salesforce, and recently Eucalyptus- an open source academic cloud. The popularity of the cloud is rapidly increasing: many computing services have now moved to the cloud and many new applications are emerging either as standalone or in orchestration with other services. As a result, the cloud architecture is dynamically scaling up to accommodate such growth. Such growth necessitates dynamic approaches for maintaining and monitoring an acceptable level of Quality of Service (QoS). From the architecture point of view, scaling up such systems does certainly introduce additional computational power, as meeting QoS requirements such as scalability, availability, reliability, real time performance, fault-tolerance, openness and security could imply the need for additional computational resources: this may, for example, entail hardware and software redundancy to be deployed in realization to some architectural mechanisms and tactics, like load balancing, replication, migration transparency, and so forth. The situation will lead to an uncontrolled growth of computational power. Such growth, if left unmanaged, is expected to contribute to the degradation of our ecosystem as we move and heavily depend on the cloud. Meanwhile, meeting QoS requirements is critical and can’t be neglected in favor of power savings. The research will propose a framework for dynamic self-optimization of the cloud architecture taking into account the tradeoffs involved in maintaining acceptable QoS with minimal power at runtime.

The proposed work is a multidisciplinary and will involve the following schools and institutes at Birmingham. These are School of Computer Science, The School of Electrical, Electronic and Computer Engineering, Institute for Energy Research and Policy, and the Department of Economics at Birmingham University. We also plan to collaborate with researchers in Cercia working on relevant themes of the proposal. The proposal will benefit from the involvement of the following named researchers, who provided their support to develop the work progarmme into a proposal:

  • Dr Rami Bahsoon(The School of Computer Science, Lecturer in Software Engineering, cloud computing; self-managed software architectures; economics-driven software engineering; application of Dynamic Data Simulations Systems to software engineering);
  • Dr Xiao-Ping Zhang (The School of Electrical, Electronic and Computer Engineering- Reader in Energy Distribution Systems and Director of the Institute for Energy Research and Policy. Dr Zhang is an expert in the application of power electronics such as FACTS - Flexible AC Transmission System and CUSTOM POWER in transmission and distribution systems; technologies for smart grids; power system economics; large scale power system optimization and planning; analysis and control of power system stability; Power quality and harmonics; integration of distributed generation -embedded generation- into electrical power systems; micro-generation and Micro-grid);
  • Dr Georgios Theodoropoulos(The School of Computer Science- Reader in Distributed Systems. Dr Theodoropoulos is an expert in the area of Dynamic Data Simulations Systems and Distributed Systems; Grid; High Performance Computing);
  • Prof Indrajit Ray (Department of Economics- Chair in Economics Theory. Prof Ray is an expert in Game Theory; General Equilibrium Theory; Social Choice Theory; Experimental Economics; and Environmental Economics).

Symbiotic blending of advanced machine learning and data mining with astronomy. Peter Tino, School of Computer Science

Large archives of astronomical data (images, spectra and catalogues) are being assembled into a publicly accessible dataset, available worldwide as part of the Virtual Observatory (VO, see www.ivoa.net). This necessitates the development of techniques that will allow fast automated classification, parameter extraction, characterisation and visualisation of multi-dimensional and multi- type datasets. Research on applications of machine learning (ML) and data mining (DM) in astronomy has recently grown in intensity and ever increasing volumes of new data of increasingly better resolution provide an ideal playground for new research developments in both scientific disciplines. In contrast to biology or finance, there typically are physics based theories behind the phenomena studied by astronomers. However, most applications of ML and DM to astronomy have so far not taken advantage of this because of an emphasis on more traditional statistical analysis tools. We believe that only by combining the underlying physics with machine intelligence on a deeper level can the true potential of ML/DM in astronomy be realised. As pilot studies of a deeper symbiosis of the two scientific fields we will concentrate on two large projects – one in the context of supervised, the other in the context of unsupervised learning.

1. Principled calibration of astronomical simulations using novel semi-parametric density estimation techniques and a new framework of multi-manifold learning. Increasing computational power enables researchers to simulate in increasing detail complex astronomical processes. Yet, how can such simulations be verified is currently an open question that needs to be urgently resolved. First, the simulations involve massive numbers of simulated elements (e.g. stars) that cannot be compared with real observations on a point by point basis. Second, the simulations are conducted in a large number of dimensions, whereas the real observations are often lower-dimensional, incomplete and noisy. We propose to calibrate simulations by constructing densities in the simulation space, projecting them onto the observation space and then performing principled model evaluation using the available observations. As a test case we will calibrate large-scale simulations of a galaxy satellite disruption. The work will be done in collaboration with Prof. Arif Babul (University of Victoria, theoretical models of galaxy interactions in their cosmological context), Dr. Mark Fardal (University of Massachusetts, numerical simulations of galaxy interaction and disruption) and Prof. Puragra Guhathakurta (University of California, optical observations of interacting galaxies).

2. Efficient learning with privileged information. In contrast to traditional ML/DM approaches, the recently introduced framework of learning with privileged information (PI) allows one to construct a learning machine using data from the `input space' X, as well as additional (privileged) data from a space Y that constitutes a `context' for X. Crucially, only data from X is available in the `test' phase (once the model is constructed and used) – data from Y serves solely as a contextual information during the model fitting. This framework can be naturally used to mine Virtual Observatory data with sparsely distributed detailed information on objects of interest. For example, when predicting red shift of a galaxy, only relatively cheap-to-obtain flux (as well as few coarse spectral) features are used. Yet, for many objects full spectral information is available and it is wasteful not to use such detailed (privileged) information in the model construction. However, the framework of learning with PI was formulated in the context of kernel-based support vector machines (PI is used for construction of the slack model) and hence cannot be used for mining vast data sets. Translating the idea of learning with PI outside support vector machines is highly non-trivial. We will develop a new boosting-based framework for learning with PI that will enable efficient construction of physics based predictive machine learning models on vast data sets. The framework will be tested on red-shift prediction and galaxy classification tasks in collaboration with Prof Ajit Kembhavi (Director, Virtual Observatory India), Dr Yogesh Wadadekar (National Centre for Radio Astrophysics, India) and Prof. Alain Omont (Institut d’Astrophysique, Paris), all of whom have been involved in original data compilation and analysis with straightforward tools.

Positron Emission Particle Tracking (PEPT) applied to multiple particle tracking in the processing of materials William Griffith, School of Metallurgy and Materials

Positron Emission Tomography (PET) is a medical physics technique, which uses a radioactive isotope that decays by release of a positron. The positron travels in the local material until it comes into contact with an electron, which results in their mutual annihilation and the production of a pair of γ-rays which are emitted back-to-back, 180° to each other. Detection of these γ-ray pairs allows the original position of the radioactive species to be established by triangulation, with the accuracy of its location determined within a few mm. In Positron Emission Tomography (PET) the technique can be used to develop a map of the concentration of a radioactive species as it is dispersed. In a related technique, Positron Emission Particle Tracking, (PEPT), a single radioactive particle can be added to a moving fluid, and its trajectory determined, also to within a few mm.

The proposed grant application is intended to fund research into the development of PEPT techniques, together with development of mathematical models of particle behaviour in fluids, so that the a better understanding of multiple particle interactions can be obtained. The outcomes will be applicable to a range of industrial processes where fluid flow and solid particles are involved. The proposed grant application is to be aimed at the development of experimental techniques using PEPT, and the development of models of particles in fluids, in in order to examine multiple particle interactions and obtain better predictive models of their behaviour.

Specifically, the proposed grant application can be viewed as consisting of three parts. The limits of resolution and accuracy of the PEPT technique will be enhanced by refining the algorithm currently used to interpret data from the technique as it now stands. This will greatly improve the ability of the PEPT technique to track multiple particles simultaneously. It will also allow better tracer particles to be used, better in the sense that they could be smaller, and of different chemical species, and therefore more representative of the types of solid particles encountered in the fluids found in the industrial processes under consideration. Secondly, general models of multiple particle behaviour in fluids will be developed using the Discrete Element Method, by collaboration of all parties. The construction of these models will be informed by the information gained by the PEPT technique, and experimentally validated by the same technique also. Thirdly, models of particle behaviour in the two specific industrial areas identified here, metallurgical processing and food processing, will be developed out of this experimental and modeling work, and will use both Positron Emission Tomography and Positron Emission Particle Tracking techniques.

The overall goals of the grant application will be; (i). better understanding of particle behaviour in liquids, (ii) accurate models of particle behaviour in liquids and, (iii). models of particle behaviour in liquids that are comprehensively validated. Successful application of the grant will allow a better understanding of multiple particle behaviour in applications relevant to numerous production processes involving materials with a wide range of properties, outside of the two specific areas identified here.

First Fellowship Award Winners

We received 6 applications to our first round of funding. All of them were of high quality and could have been funded, in the end the board decided to fund these two projects:

A vector-support machine analysis of functional recovery from stroke Glyn Humphreys, School of Psychology

The research project will examine whether advanced pattern classification procedures (using vector support machines) can be used, along with behavioural data, to predict functional recovery after stroke. Stroke affects around 150,000 people in the UK per year, and cognitive problems are both common (affecting over 70% of the patients recruited through our Stroke Association cognitive screening measure) and linked to poor long-term recovery. Being able to predict functional outcome will provide a means of optimizing the use of NHS resources, but currently there are few good prognostic models of functional outcome.
Our aim will be to combine our psychological analysis of cognitive problems following stroke with state-of-the-art data analysis techniques for assessing, from neural imaging data, the predictors of functional recovery. This will include training vector support machines to classify patients making good and poor recovery, combining data from CT with our behavioural results.

Computational modelling of ionomer glasses and glass ceramics for medical applications: Modelling the structure of the glass network and understanding the effect of cation substitutions Artemis Stamboulis, Metallurgy and Materials

Ionomer glasses are used today as fillers in glass ionomer cement restorations in dentistry. When crystallised they have the potential to form strong glass ceramics with very good mechanical properties, machinability and biocompatibility (osteointegration) for use in bone repair. The biggest market however is the dental market. The structure of the glass as well as the composition, are very important parameters for the setting in glass polyalkenoate cements as well as the crystallisation and the crystallisation mechanism of the glass ceramics.
We have currently a large amount of experimental data that would help enormously to constrain the search space for new compositions of materials in order to apply methods such as reverse Monte Carlo simulation for finding minimum energy configurations of the complex chemical compounds we are interested in. Given the complex nature of our task and the vast search configuration space, we will develop novel dedicated sampling algorithms guided by novel computational materials design strategies based on experimental data. The computational learning component will enable the sampling algorithm to accommodate prior domain knowledge, as well as experience from previous runs of the simulations. We expect that such models will be more effective than the current approaches for studying and modelling glass structures.
The following are the main objectives of the research proposal: Use the developed novel sampling-based simulation model to:
  1. Better understand the dynamics of the glass network such as configuration and conformation of the network
  2. Lay out the rules governing the structure and crystallisation mechanisms in the above glasses and glass ceramics
  3. Predict the properties and biological response of the above glasses and glass ceramics.