(opens in new tab)<\/span><\/a>.<\/p>\n\n\n\n\n\nSelected PhD Projects \u2013 2020<\/h3>\n\n\n\n
The following applications were selected for funding starting in the academic year 2020\u20132021.<\/p>\n\n\n\n\n\n
Supervisor:<\/strong> Amos Storkey, University of Edinburgh, UK
MSR Supervisor:<\/strong> Katja Hofmann<\/p>\n\n\n\nSummary: <\/strong>Deep reinforcement learning (RL) has had huge empirical success and is a major enabling technology for many applications of AI. However, recent RL algorithms still require millions of samples to obtain good performance. Since obtaining environment interactions is often costly and since challenging environments are rarely static, this inhibits many practical applications.\u202f This project will investigate ways of reducing this cost, aiming to find more sample-efficient RL algorithms. We aim for the algorithms to be deployable in realistic settings, where agents use deep networks to represent knowledge about the environment. Improving sample efficiency of RL has immediate applications to Microsoft\u2019s efforts in applying RL games. It is also likely to lead to improved performance of other systems making automated decisions.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Gavin Doherty, Trinity College, Dublin
MSR Supervisor:<\/strong> Anja Thieme<\/p>\n\n\n\nSummary: <\/strong>Health self-report or self-monitoring activities, such as mood logging, are a central part of many treatments for mental disorders. Mood logs need to be recorded regularly to be beneficial, yet people can find it difficult to stay engaged with them over time. Mood logging is an example of a health status reporting task, and while mental health is a huge issue, management and self-management of many other health conditions also involve some form of health status reporting. We propose that health status reporting systems based on speech and conversational user interfaces (CUI) have potential advantages in terms of accessibility, engagement, and disclosure. Speech as a modality can be natural and convenient for users, and may support users to disclose their feelings and help create a reporting experience that is engaging and lightweight. The act of speaking aloud might also increase the sense of unburdening associated with mood disclosure, and help provide an experience which is cathartic. The aim of this PhD is to explore the feasibility of conversational user interfaces (CUIs) as a means for supporting health self-reports of mood logging, while gathering design based insight for the development of such systems. The PhD research will establish the feasibility of this approach to mood logging, together with a detailed exploration of the design space, identification of appropriate strategies, and a carefully designed best-of-breed example of a conversational health status reporting interface. The results will inform the development of a set of design guidelines for health status CUI\u2019s. The provision of richer, natural language based health status reporting data also opens up opportunities for further research on the application of machine learning to mental health status data.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Joe Finney, Lancaster University, UK
MSR Supervisor:<\/strong> Steve Hodges<\/p>\n\n\n\nSummary: <\/strong>This project proposes the creation of new tools and processes that will enable embedded hardware devices to be successfully manufactured in low volumes \u2013 thousands of units and fewer. This is important in a world where citizen developers are becoming empowered to design new hardware solutions ranging from innovative interactive devices to IoT sensing and control systems. Recent advances in DevOps offerings from Microsoft and others have enabled these citizen developers to deploy software-only solutions at any scale, but there are no equivalent \u201chardware DevOps\u201d tools which facilitate deployment of novel hardware. By engaging with Microsoft Research and other industrial partners, this PhD project will start by identifying requirements common to the reproducible manufacture of a range of embedded hardware devices. From these hardware tooling design patterns that support low volume yet reproducible manufacturing will be proposed, built and evaluated. The ultimate aim is to create a set of freely available open tools and processes that others can use and build upon, to unlock a long tail of hardware devices.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Thomas Bohn\u00e9, University of Cambridge, UK
MSR Supervisor:<\/strong> Sean Rintel<\/p>\n\n\n\nSummary: <\/strong>Our project will be the world\u2019s first research endeavour to study how artificial intelligence (AI) influences the behaviour of humans when they interact with new forms of virtual humans. Our research is motivated by recent technical developments enabling digital representations of humans that are visually and behaviourally high-fidelity (up to indistinguishable) from real humans. The potential impact of these virtual humans on society is expected to be substantial and cannot be understood without new research. Our research team is specifically interested in Avatar-Agent-Hybrids, which can be controlled either by a human or by an AI. Avatar-Agent-Hybrids offer new research opportunities in which humans cannot distinguish if the digital representations they interact with are human or artificial agents. As this type of interaction is likely to affect especially how humans work, our project focuses on work-related situations. By developing a novel experimental setup in virtual reality (VR), we will simulate \u2013 in real-time \u2013 realistic social and work contexts in which we can study the effects of visually and behaviourally indistinguishable human avatars on humans and their social interaction. This will allow our research team to gain ground-breaking and fundamentally new cognitive and behavioural insights into the effects of AI on humans.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Diego Perez Liebana, Queen Mary University of London, UK
MSR Supervisor:<\/strong> Sam Devlin<\/p>\n\n\n\nSummary: <\/strong>The latest breakthroughs in game playing AI have primarily focused on the application of Reinforcement Learning (RL) and Statistical Forward Planning (SFP) methods, such as Monte Carlo Tree Search (MCTS) and Rolling Horizon Evolutionary Algorithms (RHEA), to games. Advances in Go, Chess, Shogi, Atari and Starcraft have been achieved by combining Deep Learning and different forms of model-free or model-based RL, including variants of MCTS. Model-based and SFP algorithms require access to an internal model of the game that allows agents to reason about the future, enabling a more data-efficient and flexible behaviour than those learnt with model-free Deep RL. However, in many real world applications and complex games (such as many created in the games industry), this model is non-existent, not available, or computationally too expensive to use. There is a prominent line of research that currently aims at automatically learning these models from interactions with the environment, showing that learning a forward model provides an important boost for action decision making. However, the scenarios normally employed in the literature are simple. Using the Malmo platform, this proposal focuses on approximating forward models in complex games by learning local interaction functions, to then investigate the use of SFP algorithms in these domains. Specifically, this research aims at (1) learning local forward models in 3D environments with partial observability; (2) implement SFP methods that would plan over said learnt models; and (3) investigate how the presence of non-stationary policies of other agents affect both model learning and SFP performance.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Matthew Foreman, Imperial College, UK
MSR Supervisor:<\/strong> James Clegg<\/p>\n\n\n\nSummary: <\/strong>Measuring birefringence in 3D is important for many applications including optical data storage and cell biology. The current state-of-the-art techniques use methods that are designed to measure thin (2D) structures but have been adapted in ad-hoc ways and combined with advanced processing to recover 3D information. Remarkably, there is an important and basic research question that still exists: is there a physically correct way to isolate and measure the birefringence of a thin 2D layer that is contained within a 3D volume, and therefore correctly reconstruct a full 3D volume? This PhD research project will answer this question.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Harish Bhaskaran, University of Oxford, UK
MSR Supervisor:<\/strong> Francesca Parmigiani<\/p>\n\n\n\nSummary: <\/strong>The use of functional materials that can accumulate information to carry out both memory and computing tasks in-situ is a growing field. Using optical integration onto silicon chips provides a unique opportunity to combine the benefits of silicon scaling and the wavelength multiplexing of optics onto a single platform. In this proposal, we will significantly expand our initial work on carrying out non-von Neumann computations (such as Vector-Matrix Multiplication directly in hardware) to a larger matrix to prove a lab scale demonstration of the potential for such hardware in real-world computing tasks. In addition, this proposal will aim to benchmark such computations against existing state-of-the-art in terms of energy and operational speed.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Ben Glocker, Imperial College London, UK
MSR Supervisor:<\/strong> Ozan Oktay<\/p>\n\n\n\nSummary: <\/strong>Being able to detect and communicate when a predictive model fails is of utmost important in system that are used for clinical decision making. In this PhD research project, novel mechanisms for failure detection and inferring causes for failure will be developed and tested on large scale population data. It is expected that the research leads to high impact innovation that is critical for deploying learned based systems and services in healthcare settings.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> James Locke, University of Cambridge, UK
MSR Supervisor:<\/strong> Andrew Phillips<\/p>\n\n\n\nSummary: <\/strong>Programming biological systems in cell populations is a key goal of synthetic biology. To achieve this, it is critical to understand how synthetic genetic circuits behave in individual cells. This is because gene expression and growth can vary substantially between individual cells, which can impact the function of synthetic gene circuits. This PhD will combine mathematical modelling, machine learning, and synthetic biology techniques to predict the population level behaviour of gene circuits from single cell gene expression data. We will use single cell time-lapse microscopy and microfluidics to characterise already constructed synthetic gene circuit components for programming self-organised cell behaviour (Aim 1). Next, we will use modelling and machine learning to design and build spatial signalling and patterning systems based on our single cell data (Aim 2). Finally, we will test our modelling predictions using custom microfluidics that allows spatial signalling between cells (Aim 3). We aim to develop a series of robust spatial patterning circuits, using the data inferred from our single cell experiments to guide our synthetic biology efforts.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Philipp Hennig, Eberhard Karls Universit\u00e4t T\u00fcbingen, Germany
MSR Supervisor:<\/strong> Cheng Zhang<\/p>\n\n\n\nSummary: <\/strong>Tree search is a classic computational task and yet still relevant for machine learning applications. Phrased as (sequential) active and reinforcement learning, it is arguably the most extreme case of \u201csmall-data AI\u201d, since it requires learning in an exponentially large decision space from linearly limited data. Recent successes in Go and related domains proved that statistically motivated best-first search strategies, guided by good heuristics, can drastically outperform the classic rule-based algorithms. Even contemporary best-first algorithms, however, are derived with local worst-case (bandit) formalisms that do not consider structure in the search domain, and only propagate a single point estimate up the tree. We propose a research project to develop a probabilistic formalism for tree search. The resulting algorithmic framework will include classic touchstones as corner cases and make clear connections to other formalisms for sequential decision making in ML. It will be able to leverage structure of the search domain through kernels and other similarity measures and use probabilistic decision theory to locally improve search efficiency. The project proposal allocates significant research time for foundational theoretical and algorithmic research, complemented by application to practical domains of relevance for MSR. A suitable PhD student has already been identified.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Stefanos Kaxiras, Uppsala University, Sweden
MSR Supervisor:<\/strong> Boris Kopf<\/p>\n\n\n\nSummary: <\/strong>This proposal addresses speculative side-channel attacks, i.e., Spectre-type attacks, at the architectural level. Protecting software from malicious attacks is already a daunting task without having to worry about fundamental guarantees at the architectural level. As a community, we were complacent that if security at the software level fails at least security at the architectural level will safeguard our most valuable secrets (e.g., secure enclaves, Intel SGX), keep adversaries contained\/confined (e.g., process and address space separation), etc. This conviction was shattered in January 2018 when speculative side-channel attacks were revealed. These attacks break software and hardware barriers that are commonly raised against adversaries. This characteristic sets these attacks apart from any past threat. It is, therefore, of utmost urgency to address them. We simply cannot reason about computer security without being able to reason about the architectural foundations of security.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Sophia Drossopoulou, Imperial College London, UK
MSR Supervisor:<\/strong> Matthew Parkinson<\/p>\n\n\n\nSummary: <\/strong>Adoption of the cloud requires trust, but software vulnerabilities can quickly erode that trust. There is a real drive in the industry to make memory safety vulnerabilities a thing of the past. Project Verona is a highly ambitious research project to make that a reality for the infrastructure we build for the cloud. This Ph.D. scholarship will apply world-class research on semantics and type systems to the design of Project Verona. This will provide the secure foundations we need for trusting the cloud.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Graziano Martello, University of Padua, Italy
MSR Supervisor:<\/strong> Sara-Jane Dunn<\/p>\n\n\n\nSummary: <\/strong>This proposed project draws from a longstanding interdisciplinary collaboration between the two named supervisors and Dr. Boyan Yordanov, which has to-date led to high impact publications in the domain of stem cell research. The project makes use of our previous work and extends the research focus to the domain of human pluripotency \u2013 the unique embryonic state in which a stem cell is poised to generate all lineages of the adult body. This is a timely project, as the stem cell field is beginning to learn more about the critical regulators of pluripotency in human, and how we can leverage the power of reprogramming and differentiation to generate cells for translational and pharmaceutical research. The aim of the proposed project is to uncover the biological program governing pluripotency in humans, and to utilise this understanding to inform reprogramming protocols, as well as our understanding of stem cell differentiation.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Amir Shaikhha, University of Oxford, UK
MSR Supervisor:<\/strong> Tom Minka<\/p>\n\n\n\nSummary: <\/strong>This proposal aims to build a unified compilation-based framework for processing of a wide range of numerical processing tasks with guaranteed efficiency. The key idea is to express such tasks in terms of a class of algorithms with strong asymptotic guarantees thanks to their nice algebraic structures. The techniques developed in this proposal will be integrated into the Infer.NET framework in order to make them applicable for all the existing applications developed using this framework, as well as new applications for automatic differentiation and abstract interpretation.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Oussama Metatla, University of Bristol, UK
MSR Supervisor:<\/strong> Cecily Morrison<\/p>\n\n\n\nSummary: <\/strong>Children born blind often have substantial social difficulties thought to stem from challenges in building up basic attentional processes that are usually developed as babies through vision. It is likely that spatialized audio could provide an alternative channel whereby visual social information is translated into spatial information to address these fundamental development challenges which have life-long consequences for social interaction and learning. This PhD would build on the Project Tokyo system and the initial findings of research done with the aim of understanding the long-term impact of the technology. The PhD is broken down into three parts: 1) Establishing appropriate longitudinal mechanisms for measuring joint attention in blind children; 2) Designing the human-AI partnership experience for long-term usage based on the initial Tokyo study; and 3) Evaluating the effects of long-term use of this technology on spatial attention in a cohort of blind children.<\/p>\n\n\n\n\n\nSupervisor:<\/strong> Tomasz Trzci\u0144ski, Warsaw University of Technology, Poland
MSR Supervisors:<\/strong> Marek Kowalski, Nate Kushman<\/p>\n\n\n\nSummary: <\/strong>High quality human-like animations are required for many different applications. In video gaming many popular games involve controlling and interacting with humanish characters. In AR and VR a core component is realistic rendering and animation of other people\u2019s avatars. In filmmaking, CGI often involves generating footage for sequences that involve the human characters in the film.The common methods used for creating and animating digital characters are both time consuming and difficult to scale. The creation process usually begins with a 3D scan of a real person\u2019s face or body performing various actions. The scan requires specialized equipment and needs to be further processed by artists. To animate the model in a realistic way an actor needs to perform the desired sequences in a motion capture studio. This effort is extremely time consuming and expensive.<\/p>\n\n\n\n\n\n\n\n\n\nA Smart Care System for Healthcare using Contextual Reinforcement<\/h3>\n\n\n\n
Supervisor:<\/strong> Mihaela van der Schaar, University of Cambridge, UK
MSR Supervisor:<\/strong> Danielle Belgrave <\/p>\n\n\n\nSummary:<\/strong> Clinicians are routinely faced with the practical challenge of integrating longitudinal, multi-modal data collected from a variety of sources (including wearables) for a given patient and then making decisions (monitoring and intervention choices) on the basis of what is learned from this data. This work is applicable to numerous areas in healthcare, including chronic diseases, mental health, elderly care etc. As the range of available information and interventions increases, the difficulty of selecting appropriate information collection and interventions for a particular<\/em> patient grows as well. As a result, the rate of misdiagnosis and mistreatment remains high. More and more data is collected for each patient and more and more data is available from past patients. The challenge is to use this data to personalize diagnosis and interventions. Current diagnosis and treatment which continues to rely on Clinical Practice Guidelines (CPGs), are geared towards the \u201clowest common denominator\u201d and are targeted toward a \u201crepresentative\u201d patient rather than toward the unique characteristics and circumstances of the current patient.<\/p>\n\n\n\n
\n\n\n\nDeep Learning for Graph-structured Data<\/h3>\n\n\n\n
Supervisor:<\/strong> Jose Miguel Hernandez-Lobato, University of Cambridge, UK
MSR Supervisor:<\/strong> Alexander Gaunt <\/p>\n\n\n\nSummary:<\/strong> We describe work to investigate the limitations of graph-based deep learning models and provide extensions that overcome these limitations. Graphs are a powerful representation of relations in interacting systems, and the ubiquity of graph-structured data in a variety of natural science and engineering domains has generated great excitement in deep learning models that capture graph structure. Typically, these approaches implement a message passing algorithm where by each node aggregates information from its incident edges to represent its environment. This aggregation is lossy and the message passing schedule (e.g. number of rounds of message exchanged) is not tailored to the complexity or embedded structure (e.g. hierarchical subgraphs) of each processed graph. The proposed work will target these limitations, yielding more performant models and providing opportunity to better interpret the model predictions using the exposed embedded structures or highly activated subgraphs.<\/p>\n\n\n\n
\n\n\n\nDesigning for Human Partnership with AI in Everyday Life: The Future of Work as a Case Study<\/h3>\n\n\n\n
Supervisor:<\/strong> Elisa Giaccardi, Delft University of Technology, The Netherlands
MSR Supervisor:<\/strong> Richard Banks <\/p>\n\n\n\nSummary:<\/strong> The proposed PhD research will study social, informal work practices \u2013 both distributed and colocated \u2013 as a case study for emergent forms of Human-AI partnership. In the future, people at work will be surrounded by AI systems, both hardware and software based. This PhD would look at how these systems should adapt to the complexities of the working environment, particularly as we shift from traditional forms of office work, to new, more informal ways of working that might include gig work and co-working spaces. This will be investigated by balancing and integrating traditional ethnographic methods and Wizard of Oz prototyping into a future-oriented approach that uses hypothetical and simulated AI systems as a means to explore Human-AI partnerships in context (see Methods of research). By means of ethnographic accounts of the future, the research will contribute an innovative design method for the design of AI systems, meant to help systems developers and designers to have empirically grounded discussions about critical factors, social norms and ethical rules, and to make decisions on what AI systems to develop.<\/p>\n\n\n\n
\n\n\n\nExploring motif-based design patterns for biological computation<\/h3>\n\n\n\n
Supervisor:<\/strong> Thomas Gorochowski, University of Bristol, UK
MSR Supervisor:<\/strong> Boyan Yordanov <\/p>\n\n\n\nSummary:<\/strong> In software engineering, design patterns provide reusable solutions to frequently encountered problems that are independent of the programming language used for implementation. In this project, we explore whether biology also exploits design patterns in the regulatory programs controlling life. Focusing on the role of small regulatory motifs that are known to be enriched in living systems, and which cluster in specific ways, we will employ formal methods to study the cellular functions that motifs support and their robustness to being used in different ways. This information will form the foundation of a computational tool where motifs and their clustering are used as design patterns to synthesise new regulatory programs. To test their generality, several programs implementing broad functionalities (e.g. logic, oscillations and pulse generation) will be verified in living cells with synthetic regulatory components not used by nature. Insight from this work will fuel new approaches for the design of our own biological programs, while also providing a deeper understanding of the way biology harnesses the computational substrate of life itself.<\/p>\n\n\n\n
\n\n\n\nIntegrating machine learning in to the IDE<\/h3>\n\n\n\n
Supervisor:<\/strong> Andrew Rice, University of Cambridge, UK
MSR Supervisor:<\/strong> Miltiadis Allamanis <\/p>\n\n\n\nSummary:<\/strong> Modern integrated development environments (IDE) offer a whole suite of tools and support to improve programmer productivity. Recent successes in the application of machine learning approaches to software development suggest that a revolution in this area is on the way. Machine learning techniques have been shown to capture subtleties of programming style in a way that has never been done with traditional analysis techniques. Recent successful results include the automated suggestion of variable names, comment generation, code summarisation, and defect detection. However, none of these techniques could be practically deployed in an IDE today. This research seeks to tackle one major challenge in this area: that of suitably designing and training a model for wide-scale use. Our vision is to deploy globally-trained models to an IDE that can then be augmented or specialised as needed. We propose three packages of work for the PhD student: 1) reimplementation and replication of key existing research results to form the basis for our experiments; 2) an investigation into the sensitivity of these techniques to their training data and the use of program-analysis techniques for understanding dataset quality; 3) development of techniques for specialising a model (or its suggestions) for style-guide differences, language versions, or library choices.<\/p>\n\n\n\n
\n\n\n\nPrivacy in Distributed and Collaborative Learning<\/h3>\n\n\n\n
Supervisor:<\/strong> Emiliano De Cristofaro, University College London, UK
MSR Supervisor:<\/strong> Olya Ohrimenko <\/p>\n\n\n\nSummary:<\/strong> Distributed and collaborative machine learning are increasingly being deployed in the wild. Because the training data either never leaves the participants\u2019 machines, or is never exposed to untrusted parties, these techniques are a good fit for scenarios where data is sensitive, and the participants want to construct a joint model without disclosing their datasets. However, since model updates\/outputs are still indirectly based on the training data, we need to measure whether, and how much, we can prevent the leakage of unintended information about the participants\u2019 training data; and, if so, what can be done to mitigate such leakage. Concretely, we plan to look for leaky inferences in real-world settings, investigate defenses, including training regularizers, such as Dropout or Weight Normalization, user-level differential privacy, adversarial learning techniques, as well as relying on black-box trusted servers. In particular, with respect to the latter, and overall centralized machine learning, we plan to study its trade-offs compared to distributed learning based on communication\/computation overhead incurred on end-users and privacy\/security guarantees it offers.<\/p>\n\n\n\n
\n\n\n\nProgramming biological systems by reverse-engineering reaction-diffusion model<\/h3>\n\n\n\n
Supervisor:<\/strong> Attila Csikasz-Nagy, Kings College London, UK
MSR Supervisor:<\/strong> Neil Dalchau <\/p>\n\n\n\nSummary:<\/strong> As genetic engineering becomes more precise, the need for high-throughput automated biological scienti\ufb01c discovery increases. The conceptual gap between computer science and biology is drawing to a close: biological processes can be mapped to algorithms and we are beginning to leverage the asynchronous logic of biochemical reactions to design soft matter computing devices. The overarching aim of this project is two-fold: to develop automated and analytical tools for the design of biochemical reaction networks in synthetic and systems biology settings; to apply these tools to the control of patterns for spatial biochemical computation.<\/p>\n\n\n\n
\n\n\n\nReinforcement Learning for Enabling Next Generation Human-Machine Partnerships<\/h3>\n\n\n\n
Supervisor:<\/strong> Adish Singla, Max Planck Institute for Software Systems, Germany
MSR Supervisor:<\/strong> Sam Devlin <\/p>\n\n\n\nSummary:<\/strong> The ultimate goal of AI systems is to support people in achieving their goals more efficiently. While recent advances in AI have led to a remarkable performance of machines in challenging tasks, e.g., in image\/speech perception and playing games like Go, these feats have largely been limited to well-specified tasks with known dynamics and predictable outcomes. These limitations can be addressed by designing AI systems that emerge from the complementary abilities of humans and machines by enabling close partnerships between them. For instance, in autonomous driving, this partnership could manifest in the form of an AI auto-pilot handing over control to the human driver in safety-critical situations. To enable this partnership, this project will focus on developing novel reinforcement learning (RL) approaches that effectively and efficiently learn with-and-from people in complex real-world environments. More specifically, we tackle the following fundamental research questions: (i) Given potential differences in perception and behavioral biases between machine and human, how can we design robust multi-agent RL algorithms? (ii) Given that a human could adapt its behavior in the presence of an AI agent, how can we design reactive RL algorithms for enabling long-term human-AI collaborations? (iii) How can we empower a human to steer the behavior of an AI agent, for instance by teaching interactions, and how to make this teaching process more effective? By answering these questions, the project\u2019s mission is to enable next generation human-machine partnerships for the benefit of people and society.<\/p>\n\n\n\n\n\nAI for Teams: The Future of Assisted Collaborative Work<\/h3>\n\n\n\n
Supervisor:<\/strong> Alex Taylor, City, University of London, UK
MSR Supervisor:<\/strong> Sean Rintel <\/p>\n\n\n\nSummary:<\/strong> The proposed PhD research will study the uptake of AI in team-working tools, with a particular emphasis on shaping the future design of this new genre of collaborative work. A qualitative methodology, using predominantly an ethnographic orientation, will be applied to investigate contemporary modes of collaborative work and specifically how AI-based tools and services are being incorporated into work practices and knowledge-based activities. As with long-established workplace studies, the detailed and in-depth results of this orientation and analysis will be used to develop design thinking in this domain. To best exploit this, the PhD research will run alongside and intersect with concurrent work at Microsoft Research and in the Human Experience and Design group focused on teams and the integration of AI.<\/p>\n\n\n\n
\n\n\n\nCompartmentalized RNA Computing using Protein-based Microcompartments<\/h3>\n\n\n\n
Supervisor:<\/strong> Tom de Greef, Eindhoven University of Technology, Netherlands
MSR Supervisor:<\/strong> Andrew Phillips <\/p>\n\n\n\nSummary:<\/strong> DNA strand-displacement has been widely used for the design of molecular circuits, motors, and sensors. Typically, DNA-based molecular computations based on strand-displacement are executed under non-compartmentalized conditions where all DNA gates are, in principle, able to interact with each other. Recently, we developed an experimental platform that allows compartmentalization of DNA gates into protein-based microcompartments. In addition, we have shown that addition of single-stranded DNAs to the outside of the microcompartments is able to trigger a DNA-based strand displacement reaction from a localized DNA gate, which in turn is able to trigger a second strand-displacement in a different microcompartment. Here we propose to further expand the scope of our experimental platform by using microRNA (miRNA) as chemical inputs. In addition, we will investigate the performance of our compartmentalized RNA and DNA circuits in biological relevant media such as bovine serum. Successful completion of the project will pave the way for distributed DNA computing using both RNA and DNA inputs in biological relevant conditions, with potential applications in high-precision medical diagnostics.<\/p>\n\n\n\n
\n\n\n\nEnabling Repair in Goal-Directed Natural Language Personal Assistants<\/h3>\n\n\n\n
Supervisor:<\/strong> Joel Fischer, University of Nottingham, UK
MSR Supervisor:<\/strong> Nate Kushman <\/p>\n\n\n\nSummary:<\/strong> Recent advances in speech recognition have brought closer the long-standing vision of the ubiquitous intelligent personal assistant. This has become particularly relevant in the context of smartphones and in-home devices where natural language interaction has the potential to significantly improve the user experience. While already extremely useful, these agents fall far short of the future we have imagined. In general, the speech recognition can successfully transcribe the uttered words, but quite often the agent still does not infer the desired intent correctly. This is particularly evident in the case of \u2018trouble and repair\u2019, where the agent\u2019s response does not correspond to the user\u2019s expectation and requires \u2018repair\u2019 through subsequent repetition and rephrasing (Porcheron et al., 2017a). Our own prior research shows that this is not an exception, but the norm; in our study of conversational interaction with the Alexa agent, more than half of 800+ requests we recorded \u2018in the wild\u2019 were unsuccessful initially and occasioned the user to follow up with further requests (Porcheron et al., forthcoming). Repair may often involve some form of \u2018guesswork\u2019 as to what went wrong, as the agent does not generally account for what may have gone wrong (e.g., explain, describe, or offer an interpretation of some form). Arguably this amounts to an example of what has been criticised as \u2018black box\u2019 AI, which has recently led to significant funding calls for research proposals for \u201cexplainable AI\u201d. In this PhD we propose the exploration of methods to facilitate repair in goal-directed natural language dialog between a human and a AI assistant, contributing towards the higher level aim of making AI more transparent, or intelligible, to user(s). This PhD aims to investigate three general different forms of repair: 1. The AI agent\u2019s response does not match the user\u2019s intent, the human would like to course correct towards the desired outcome. To facilitate repair the PhD will explore the potential benefits of making the AI agent\u2019s \u2018understanding\u2019 available to the user, and methods to direct the user towards outcomes which are both feasible for the agent and at least more closely aligned with the user\u2019s intent (i.e., a mutual configuration). 2. The user has made an utterance for which the machine cannot confidently infer the intent. To facilitate repair, the PhD will explore methods to seek clarification from the user. 3. The human has made a mistake and needs to correct themselves. The PhD will explore avenues for the user to initiate self-repair.<\/p>\n\n\n\n
\n\n\n\nIntrinsically Motivated Exploration for Lifelong Deep Reinforcement Learning of Multiple Tasks<\/h3>\n\n\n\n
Supervisor:<\/strong> Pierre-Yves Oudeyer, Inria, France
MSR Supervisor:<\/strong> Katja Hoffmann <\/p>\n\n\n\nSummary:<\/strong> This project aims to develop autonomous lifelong machine learning techniques that enable virtual or physical intelligent agents to acquire large repertoires of tasks in open uncertain environments. This is key for developing intelligent agents that need to continuously explore and adapt interaction skills to new or changing tasks, environments, people to interact with, and preferences of others. The approach will leverage recent advances in curiosity-driven developmental learning (also called intrinsically motivated learning) to drive exploration in a Deep reinforcement learning framework. They will be evaluated on benchmarks involving the Microsoft Malmo\/Minecraft video game platform, for autonomous learning of intelligent non-player characters controllers that can adapt creatively to changing environments.<\/p>\n\n\n\n
\n\n\n\nMachine Learning for Program Analysis and Defect Prediction<\/h3>\n\n\n\n
Supervisor:<\/strong> Philipp R\u00fcmmer, Uppsala University, Sweden
MSR Supervisor:<\/strong> Marc Brockschmidt <\/p>\n\n\n\nSummary:<\/strong> This research proposes harnessing recent dramatic advances in machine learning for automatic program analysis and defect prediction. The project will break new ground in the way machine learning models process programs, advancing from a mostly syntactic treatment of source code to methods that take semantics and existing program analysis techniques into account. This will enable flexible and sophisticated high-level guidance: of developers and test engineers towards likely cases of program defects, and of program verification tools towards complex program annotations and invariants.<\/p>\n\n\n\n
\n\n\n\nMechanistic Model Misspecification<\/h3>\n\n\n\n
Supervisor:<\/strong> Richard Wilkinson, University of Sheffield, UK
MSR Supervisor:<\/strong> Ted Meeds <\/p>\n\n\n\nSummary:<\/strong> Mechanistic or simulation-based models are used in scientific research to understand complex natural phenomena. A mechanistic model can take the form of ordinary\/partial\/stochastic differential equations (O\/P\/SDEs) and can be rigid in form but have the benefit to the scientist of having interpretable and testable parameter settings. In part due to the inflexibility of the model forms, misspecification of the model can lead to computationally expensive inference procedures, and more importantly, misleading conclusions, whereby the parameter estimations are confidently incorrect. There is increasing evidence that the inference framework called approximate Bayesian computation (ABC) is more robust to model misspecification than other simulation-based parameter estimation techniques. We propose to study the mathematical and statistical properties of this robustness, and explore improvements over current model misspecification approaches. The research will be pragmatic, embedding the theory with practical examples (where domain knowledge is understood, and hence misspecification can be detected), including using semi-mechanistic models that are used in the public health domain.<\/p>\n\n\n\n
\n\n\n\nOptical Fabric for Zero-\u00ad\u2010latency memory disaggregation at Pb\/s cluster-\u00ad\u2010 scale network<\/h3>\n\n\n\n
Supervisor:<\/strong> Georgios Zervas, University College London, UK
MSR Supervisor:<\/strong> Hitesh Ballani <\/p>\n\n\n\nSummary:<\/strong> Data centres have been historically based on a server-centric approach with fixed amounts of processor and directly attached memory resources within the boundary of a mainboard tray. The mismatch between fixed proportionalities and diverse set of workloads can lead to substantially under-utilized resources (some cases even below 40%) that account for 85% of the total data centre cost. The project aims to explore memory disaggregation at Rack and Cluster level and identify the scalability limits both in terms of number of end-points, network capacity per CPU\/Memory, physical distance and the associated penalties to processing power, sustained memory bandwidth etc. It will decouple CPU from memory and use an all-optical interconnect network using scalable on-board transceivers and optical switches to minimize and offer deterministic round-trip latency, maximize throughput and bandwidth density at the lowest cost and power consumption.<\/p>\n\n\n\n
\n\n\n\nOptics for the Cloud: Cost-effective assembly for Hybrid Integrated Photonic Switches<\/h3>\n\n\n\n
Supervisor:<\/strong> Kevin Williams, TU Eindhoven, Netherlands
MSR Supervisor:<\/strong> Hitesh Ballani <\/p>\n\n\n\nSummary:<\/strong> Optical switching using photonic integrated circuits holds the promise of low-latency, data transparent routing at the packet time scale. However, the assembly of integrated photonic switches involves the attachment of tens of optical fibers with deep-sub-micron-precision and this is now considered to be one of the greatest barriers to deploying switch technology. In this work, we propose to research assembly methods which relax the precision requirements. This is inherently scalable in terms of capacity and connectivity and provides a route to System-on-Chip integration with electronics.<\/p>\n\n\n\n
\n\n\n\nPixelated Spatial Light Modulator for Phase and Polarisation Modulation of Light (PPSLM)<\/h3>\n\n\n\n
Supervisor:<\/strong> Daping Chu, University of Cambridge, UK
MSR Supervisor:<\/strong> Andreas Georgiou <\/p>\n\n\n\nSummary:<\/strong> Phase holograms have been used to split the laser beam into many spots thus increasing the write speed of the optical data storage system on glass. However, the throughput and hence the costs are affected by the lack of a physical light engine which can introduce both phase modulation (to split the beam) and polarization modulation (to encode data) at the same time. The objective of this work is to develop the technology to allow polarization and phase modulation of light using a single element. With a single device that can both, slit the beam and introduce polarization modulation, it is possible to accelerate significantly the speed of data write, reduce the amount of energy required and reduce cost, for an ultimate high-density fast-access and low-cost data storage for the cloud.<\/p>\n\n\n\n
\n\n\n\nSecurely and Efficiently Supporting Managed Language Runtimes in Confidential Computing<\/h3>\n\n\n\n
Supervisor:<\/strong> Peter Pietzuch, Imperial College London, UK
MSR Supervisor:<\/strong> Manuel Costa <\/p>\n\n\n\nSummary:<\/strong> In domains such as healthcare and finance in which regulatory and competitive reasons prohibit complete trust in cloud providers, organisations are left with limited solutions for protecting against cloud security threats. A new direction is to use hardware security features in moderns CPUs in order to protect computation in otherwise untrusted cloud environments. Trusted execution environments (TEEs) such as Intel\u2019s recently launched Software Guard Extensions (SGX) introduce the concept of secure enclaves. Yet, before TEEs will be widely used to protect real-world data-intensive cloud applications, we observe two open challenges: (1) how can we protect against vulnerabilities inside the TEE itself?; and (2) how can we support high-level programming languages inside TEEs effectively? We plan to address these challenges by exploring two complementary research directions: (a) to protect against security vulnerabilities in complex TEE implementations, we will develop new compiler-based techniques for implementing protection boundaries inside of secure enclaves. By compartmentalising TEE implementations, we can contain the impact of security vulnerabilities within a secure enclave following a privilege separation policy; (b) to support complex language runtimes inside TEEs, we will employ functionality from library OSs and investigate the security and performance implications. In particular, we will explore how common functions performed by managed runtimes, such as just-in-time (JIT) compilation and garbage collection can be supported effectively and securely.<\/p>\n\n\n\n
\n\n\n\nStatistical Learning and Adaptive Observation in Clinical Prediction: Methodology and Applications<\/h3>\n\n\n\n
Supervisor:<\/strong> Glen Philip Martin, University of Manchester, UK
MSR Supervisor:<\/strong> Danielle Belgrave <\/p>\n\n\n\nSummary:<\/strong> Limited resources restrict how frequently healthcare professionals can observe patients and\/or monitor information collated in information systems. For example, within hospital acute care admissions, vital signs are collected at 4-hourly intervals to monitor escalation\/ de-escalation of treatment. While there is scope to adapt an observation process to target high-frequency observation in patients at high-risk of adverse outcome, there is currently no systematic and automated mechanism in which to inform this process. Therefore, this PhD project aims to integrate clinical prediction models (CPMs) with adaptive observation by developing methodology in this space, which will be motivated by two real-world exemplars (acute hospital care and cancer early detection). Key objectives of this PhD will include: (i) adapting\/developing probabilistic modelling\/ machine learning frameworks in the context of adaptive observation, (ii) investigate how to maintain predictive accuracy at the extremes of a predictive distribution, and (iii) examine the clinical and statistical utility of incorporating adaptive observation within prediction models. Potential modelling strategies could include graphical models, mixed effects models and Gaussian processes. Each modelling strategy will be compared and tested through extensive simulation studies, and applied to the two clinical exemplars and associated data streams. This project has potential to improve the targeting of resources in near real-time within clinical workflows.<\/p>\n\n\n\n
\n\n\n\nReinforcement Learning for Enabling Next Generation Human-Machine Partnerships<\/h3>\n\n\n\n
Supervisor:<\/strong> Adish Singla, Max Planck Institute for Software Systems, Germany
MSR Supervisor:<\/strong> Sam Devlin <\/p>\n\n\n\nSummary:<\/strong> The ultimate goal of AI systems is to support people in achieving their goals more efficiently. While recent advances in AI have led to a remarkable performance of machines in challenging tasks, e.g., in image\/speech perception and playing games like Go, these feats have largely been limited to well-specified tasks with known dynamics and predictable outcomes. These limitations can be addressed by designing AI systems that emerge from the complementary abilities of humans and machines by enabling close partnerships between them. For instance, in autonomous driving, this partnership could manifest in the form of an AI auto-pilot handing over control to the human driver in safety-critical situations. To enable this partnership, this project will focus on developing novel reinforcement learning (RL) approaches that effectively and efficiently learn with-and-from people in complex real-world environments. More specifically, we tackle the following fundamental research questions: (i) Given potential differences in perception and behavioral biases between machine and human, how can we design robust multi-agent RL algorithms? (ii) Given that a human could adapt its behavior in the presence of an AI agent, how can we design reactive RL algorithms for enabling long-term human-AI collaborations? (iii) How can we empower a human to steer the behavior of an AI agent, for instance by teaching interactions, and how to make this teaching process more effective? By answering these questions, the project\u2019s mission is to enable next generation human-machine partnerships for the benefit of people and society.<\/p>\n\n\n\n
\n\n\n\nVisual Fast Mapping<\/h3>\n\n\n\n
Supervisor:<\/strong> Richard Turner, University of Cambridge, UK
MSR Supervisor:<\/strong> Aditya Nori <\/p>\n\n\n\nSummary:<\/strong> There has been considerable progress since 2011 on the classification of objects, textures and scenes in images. This is in substantial measure thanks to developments in deep learning and their widespread adoption and development in machine vision. It is notable that machine performance appears to differ markedly from human performance at object classification tasks in requiring big data for training. Humans however can learn to recognise new categories from data sets with remarkably few labels. This doctoral project explores possibilities for machine learning that may substantially increase the training efficiency of machine classification of images. The ideas will be tested on data sets of medical images and scenes for autonomous systems. The research is expected to impact both medical imaging and the way people interact with and train AI systems.<\/p>\n\n\n\nJoint Initiative with Informatics with University of Edinburgh<\/em><\/p>\n\n\n\n
\n\n\n\nHigh-Level Synthesis of Neural Networks for FPGAs with LIFT<\/h3>\n\n\n\n
Supervisor:<\/strong> Christophe Dubach
MSR Supervisor:<\/strong> Dimitrios Vytiniotis <\/p>\n\n\n\nSummary:<\/strong> Machine-learning applications are becoming pervasive throughout our entire society. They are already used extensively in areas such as machine translation and business data analytic and are set to revolutionise our world with applications such as self-driving cars. This has become possible thanks to the massive amount of data available for training coupled with the development of powerful parallel hardware. However, writing efficient parallel implementation for these algorithms remains a challenge for the non-experts. The presence of parallel accelerators such as GPUs (Graphic Processing Units) or FPGAs (Field-Programmable Gate Arrays) means that software has to be specifically written for these devices. Programmers have to use different programming models and often need to fine-tune their code for the special characteristics of the targeted hardware. This expensive and time-consuming process needs to be repeated every time new hardware emerge or even when the software stack is updated. To enable machine-learning expert to unlock the potential of future systems, we need to focus on new software programming model that abstract away most of the hardware details. In this project, we propose to build upon our existing LIFT project, an Open Source language and compiler initially developed in my group. LIFT combines a high-level functional data parallel language with a system of rewrite rules which encodes algorithmic and hardware-specific optimisation choices. An applications written in LIFT is able to take advantage of parallel accelerators available in the systems, transparently from the user. This proposal is about augmenting LIFT with the ability to express and optimise machine-learning algorithms and exploit effectively FPGA hardware.<\/p>\n\n\n\nJoint Initiative with University College London<\/em><\/p>\n\n\n\n
\n\n\n\nComputer Vision on the Edge<\/h3>\n\n\n\n
Supervisor:<\/strong> Gabriel Brostow