2012 PhD Scholars

Alexander KoppelhuberAlexander Koppelhuber
Johannes Kepler University, Austria

Research title: LumiConSense

Supervisor: Oliver Bimber
Microsoft Research supervisors: Shahram Izadi, Otmar Hilliges

Research summary: We present a first attempt towards a fully transparent, flexible, scalable, and disposable image sensor. Our approach is based on thin-film luminescent concentrator waveguides. These are polymer films doped with fluorescent dyes that absorb light of a specific wavelength, re-emit it at a longer wavelength, and transport it by total internal reflection towards the edges of the film. They are normally used for increasing the efficiency of solar cells. By cutting the edges with a specific pattern, we force the light-transport within the film into a two-dimensional light-field. This enables the reconstruction of images that are focused on the film. Following initial simulations, we want to gain a deeper understanding in the physics and mathematics of our imaging approach, which possibly leads to practical software and hardware prototypes that enable the implementation of novel interaction and sensing applications.

Anastasis GeorgoulasAnastasis Georgoulas
University of Edinburgh, UK

Research title: Machine Learning Methods for Formal Dynamical Systems: a Systems Biology Case Study

Supervisors: Jane Hillston / Guido Sanguinetti
Microsoft Research supervisors: Luca Cardelli, Andrew Phillips

Research summary: Techniques from computer science are having a profound effect on computational science. In the context of systems biology and representation of intracellular processes, there have previously been two powerful but distinct approaches. One, based on formal model description techniques, has developed languages and associated analysis techniques to capture the (global) dynamic behaviour of biochemical processes. The other relies on more conventional differential/difference equation representation of systems but uses advanced machine learning techniques to incorporate observations and uncertainty into representations of (local) behaviour which can be verified experimentally. The objective of this project is to investigate the amalgamation of these techniques to design a formal modelling language that incorporates observations and uncertainty, and inference algorithms to allow the use of this additional information to improve the interrogation of behaviour using model checking algorithms.

Andrey RodchenkoAndrey Rodchenko
University of Manchester, UK

Research title: Virtualization and High-Productivity for Many-Cores

Supervisor: Mikel Lujan
Microsoft Research supervisor: Tim Harris

Research summary: The goal of this project is to investigate how to implement and understand the trade-offs of managed runtime environments for future computer systems, where each chip can integrate thousands of processing cores.

Argyrios DeligkasArgyrios Deligkas
University of Liverpool, UK

Research title: Automated Design of Revenue-Maximizing Ad Auctions

Supervisor: Mingyu Guo
Microsoft Research supervisors: Yoram Bachrach, Peter Key

Research summary: Similar to combinatorial auctions, an Ad auction is a set of rules that allocate different items (impressions of different types) to different bidders (advertisers). What differentiates Ad auction design from (classic) combinatorial auction design is that, as pointed out [in Emek et al. 2011], there exists information asymmetry in Ad auctions: the website knows the types of the impressions, but the advertisers do not have this information. It is possible to exploit this information asymmetry to achieve higher revenue, for example, by hiding the gender information of a male impression, and sell it as a unisex impression.

The proposed project will focus on automated design of revenue-maximizing Ad auctions that exploit information asymmetry. We hope to answer two questions:

  1. How can the website determine what information to hide?
  2. Instead of hiding, can website sell information for profit?

Arman Idani
University of Cambridge, UK

Research title: Passive personality assessment: a psychometric and machine learning approach

Supervisor: Professor John Rust
Microsoft Research supervisor: Pushmeet Kohli

Research summary: The goal of this project is to explore the potential of personality assessment methods that are based on records of individual behaviour and preferences recorded in the online environment. Such inferred personality could be used to customize the user’s experience in the context of web-searching, shopping, and the behaviour of their computers and systems.
We aim to:

  • Measure the personality of groups and individuals “passively” and non-intrusively
  • Enhance personality theory in the light of the results of this study
  • Build models linking behaviour and preferences to personality

Ben SpencerBen Spencer
University of Oxford, UK

Research title: Inferring From Integrity Constraints Through JavaScript Analysis

Supervisor: Michael Benedikt
Microsoft Research supervisor: Matthew Parkinson

Research summary: In this project we will deal with improving the process of getting data from the “hidden web”. We will perform static analysis of the JavaScript code that is attached to web forms to discover their access restrictions and integrity constraints. These in turn can be used to optimize search tools that explore hidden web content—by avoiding useless accesses that violate the constraints or restrictions. The proposal will extend the state-of-the-art in both web information management and static analysis.

Fabian NagelFabian Nagel
University of Edinburgh, UK 

Research title: Holistic Evaluation in LINQ

Supervisor: Stratis Viglas
Microsoft Research supervisor: Gavin Bierman

Research summary: This proposal argues for the use of just-in-time compilation of LINQ queries to native code for in-memory databases. In particular, the key idea is to apply recent techniques from SQL code generation in the context of LINQ and the .NET runtime. This will result in having yet another way to evaluate LINQ queries, but one that has the potential of better exploiting the hardware and software capabilities of the underlying platform. This work will improve LINQ by (a) making the implementation of LINQ for .NET objects a lot more efficient by using type-specific high-performing, and hardware-optimised algorithms as opposed to the inefficient generic algorithms based on suboptimal nested iterations that are currently used; and (b) providing LINQ with performance that surpasses that of established relational database technology for in-memory collections.

Fiana RaiberFiana Raiber
Technion, Israel

Research title: Content-Based Relevance Estimation on the Web

Supervisor: Oren Kurland
Microsoft Research supervisors: Filip Radlinski, Milad Shokouhi

Research summary: Search over the web is a difficult challenge due to the noisy and adversarial nature of documents' content, among other reasons. Web-search retrieval approaches address this challenge by utilizing non-content-based relevance indicators (for example, those based on hyperlink and click-based information) and detecting documents that are considered, in a query-independent fashion, of very low quality (for example, spam). We plan to devise novel content-based relevance estimation approaches that address the noisy and adversarial nature of the web. As a document is deemed relevant to the information need expressed by a query if its content satisfies the need, improved content-based relevance estimation can potentially help to significantly improve overall relevance estimation.

Jelte MenseJelte Mense
University of Edinburgh, UK

Research title: Developing Novel Computational Methods to Describe and Predict Human Behaviour in Earth System Models

Supervisor: Paul Palmer
Microsoft Research supervisor: Drew Purves

Research summary: We describe a PhD project that will fundamentally improve understanding of how humans will respond to changing climate and associated environmental factors. We present two interrelated projects:

  1. We will develop a model of the relationship between the changing climate and conflict, including demographic transitions, and how it is affected by the outbreak and spread of disease
  2. We will develop a model of climate-related migration, borrowing ideas from behavioural ecology, to look at how racial tension and bounded rationality might affect how communities eventually migrate.

For both projects, an emphasis will be on these predictive models reproducing observed socio-economic metrics, largely provided by the United Nations, so that we develop confidence before we apply them to future climate scenarios.

Jinli HuJinli Hu
University of Edinburgh, UK

Research title: Machine Learning Markets

Supervisor: Amos Storkey
Microsoft Research supervisors: Peter Key, Thore Graepel

Research summary: We propose to develop market systems for solving large scale machine learning (ML) problems through Machine Learning Markets (Storkey 2011). We will investigate different mechanism design procedures, and processes for selling derived information. This will involve generating techniques for developing and building combinatorial prediction markets, and developing improved mechanisms for markets and auctions via the use of machine learning techniques. We will look at applications in the area of learning the conditional probabilities associated with other human markets or the actions of human agents.

Johannes MeyerholtJohannes Meyerholt
Max-Planck Institute for Biogeochemistry, Germany

Research title: Systematic Assessment of Uncertainty in Couples Carbon-Nitrogen Cycle Models and their Climate Feedbacks

Supervisor: Sönke Zaehle
Microsoft Research supervisor: Matthew Smith

Research summary: The new generation of land carbon-nitrogen-cycle models show that nitrogen feedbacks attenuate the responses of the carbon cycle to perturbations, thereby affecting long-term projections of climate change. The magnitude of this effect is very different between the models, leading to considerable uncertainty in projected rates of climate change. This project seeks to better understand and quantify this uncertainty by systematically assessing alternative model components in a common framework. Key observations of global carbon-nitrogen cycling will be used to evaluate competing process formulations. The thoroughly examined set of model components, linked in a common global modelling framework, will be used to make ensemble projections of the effects of future global change on terrestrial feedbacks to the climate system, systematically assessing uncertainty in these projections stemming from uncertainty in both parametric and process-formulation of global carbon-nitrogen cycle modeling.

Laura Parshotam
University College London, UK

Research title: Dynamic Modelling of HIV Recognition by the Immune System

Supervisor: Peter Coveney
Microsoft Research supervisor: Neil Dalchau

Research summary: The immune system is one of the most complex biological sub-systems within a single individual, involving the organised interaction of a vast array of molecular species both intra- and extra-cellularly. A fundamental component of the immune system is the ability of cytotoxic T lymphocytes (CTLs) to recognise and then destroy virus-infected or cancerous cells thus preventing disease progression. The overall aim of this proposal is to develop an accurate theoretical framework that combines models of the HIV virus lifecycle and peptide processing and presentation, in order to predict the dynamics of the CTL response in the host.

Miltiadis AllamanisMiltiadis Allamanis
University of Edinburgh, UK

Research title: Statistical Language Processing for Programming Language Text

Supervisor: Charles Sutton
Microsoft Research supervisors: Andy Gordon, Thore Graepel

Research summary: Complex software systems involve many components and many external libraries, which create a large demand on programmer time and attention. In this project, we envision a new class of development tool, called data-driven development tools, to ease this burden. The idea is to start with a massive corpus of code from other projects (for example, of 1 billion lines of code) and apply tools from machine learning and natural language processing (NLP) to find syntactic patterns that programmers use often. We will do this by using an idea from NLP called a statistical language model, which is simply a probability distribution over strings, for example, over all possible C# files. The main goal of the studentship will be to build a statistical language model for programming language text. Doing this will require new technology on the NLP side as well. This would enable many applications, such as a syntax-based IntelliSense, which could recommend entire code snippets to a developer.

Steven WoodhouseSteven Woodhouse
University of Cambridge, UK

Research title: Development of an Executable Model Encapsulating Blood Cell Development from Pluripotent Embryonic Stem Cells

Supervisor: Berthold Gottgens
Microsoft Research supervisor: Jasmin Fisher

Research summary: Blood cell development has long stood as a paradigm for stem cell and cancer research. This proposal is designed to explore the potential of computational modelling to advance our understanding of normal and malignant blood cells, based on the following overall hypothesis:

Executable models of complex biological systems such as blood development can encapsulate current experimental knowledge as well as provide a powerful platform for hypothesis generation to investigate the biology of both normal and malignant blood cells.

Expected outcomes of this research include:

  1. An executable model for blood development, to form the basis for similar models for other organs.
  2. New hypotheses about the consequences of leukaemia-associated mutations, with the possibility to model counter-balancing interventions as candidate therapies.
  3. Further development of software tools for modelling and analysis of biological systems in the Fisher lab at Microsoft Research in Cambridge.

Thomas DykesThomas Dykes
Northumbria University, UK

Research title: Supporting a "Sense of Home" in Care Homes: an Exploration of Digital Design with People Living with Dementia

Supervisor: Jayne Wallace
Microsoft Research supervisors: Tim Regan, Siân Lindley

Research summary: Dementia and the context of life for people living with dementia has become an increasingly important topic in human-computer interaction (HCI) and design over recent years. Dementia has a profound effect globally and on each individual living with the disease. This PhD project seeks to explore creative ways to support a sense of home for people living with dementia in care homes through the creation of innovative digital artefacts.

A creative methodology will be developed that has the person with dementia at the heart of the research and design process. This project will add a novel angle to current work in the fields of HCI and design by centring on the reconstruction of home, on residents’ agentive and creative power, and about considerations of the care home as a new phase of life.

Tomasz KuchtaTomasz Kuchta
Imperial College London, UK

Research title: Incremental and Adaptive Symbolic Execution

Supervisor: Cristian Cadar
Microsoft Research supervisor: Miguel Castro

Research summary: Symbolic execution is a software testing technique that has gained attention in recent years, due to its ability to systematically explore paths through a program and find deep errors on these paths. Symbolic execution has been successfully applied to a variety of software, but still faces important scalability challenges, such as navigating through the huge execution space of real applications, and handling expensive constraint solving queries.

In this project, we plan to address some of these challenges by focusing symbolic execution on code changes (such as program patches) by devising incremental and adaptive symbolic execution techniques that can reuse the result of previous analyses as well as dynamically react to changes in the complexity of the analysis. Applications of these techniques could include high-coverage testing of code changes, reasoning about differences between a patched and an unpatched version of a program, and synthesizing various types of code fragments.

Vu Khac KyVu Khac Ky
École Polytechnique, France

Research title: Efficient Approximations for Fast Simulations: Application to Building Designs

Supervisor: Leo Liberti
Microsoft Research supervisor: Youssef Hamadi

Research summary: Smart buildings integrate architecture, construction, technology, and energy systems; they make use of building automation, safety and telecommunication devices, and they are managed automatically or semi-automatically on the basis of local information provided by a sensor network. The functioning of such a complex system necessarily depends on several tunable parameters, with respect to which the whole system can optimized as concerns several objectives (cost, energy efficiency, ambience comfort, and so on). For any given parameter value, system performance can only be evaluated by a computationally costly simulation procedure. The object of this PhD thesis is to devise new methodologies for optimizing smart building systems under such computational constraints.