The world is currently experiencing an unprecedented era of booming proliferation of machine learning (ML) and artificial intelligence (AI). Undoubtedly, the determining reason behind this rapidly evolving adoption of ML/AI is the embrace of deep neural networks (DNNs). Neural networks had been around for decades, but the advent of faster processing in the form of GPUs and storage enabling huge amounts of "big data" allowed for the training of deeper networks which showed startling performance increases on a variety of tasks in a variety of disciplines. However, the limitations of deep learning are becoming increasingly evident. Despite deep neural networks performing exceptionally well on a range of metrics, they have also been shown to be vulnerable to adversarial examples. This was first demonstrated in the field of computer vision-certain images are classified incorrectly (often with high confidence), despite there being a minimal perceptual difference with correctly classified inputs. Adversarial examples have been found in many other applications of deep learning, such as speech understanding, models of code etc. The ease with which these adversarial examples can be found raises doubts about deep neural networks being used in safety-critical applications such as autonomous vehicles or medical diagnosis since the networks could inexplicably classify a natural input incorrectly although it is almost identical to examples it has classified correctly before. Moreover, it allows for the possibility of malicious agents attacking systems that use neural networks, strikingly, Tencent Keen Security Lab recently demonstrated that the neural network underlying Tesla Autopilot can be fooled by an adversarially crafted marker on the ground into swerving into the opposite lane. The Fellowship will create a new Centre of Excellence at Oxford aiming to make deep learning reliable, robust and deployable, creating a new capability within the UK's AI/ML research landscape. The solution will involve developing fundamental algorithms to make the training more robust, together with algorithms to give an accurate uncertainty calculation for the deep networks estimates. However it is important that the solution also takes into account efficiency. As systems become deployed in the real world in some cases they will be exposed to an ever changing data stream. For instance, as of May 2019, 500 hours of video data is uploaded to YouTube every minute [2], 2.5 quintillion bytes of data are produced by humans every day. In the last two years alone, and most astonishingly, 90 percent of the world's data has been created. The challenge now, however, is to train the (almost) trillion parameters networks with quintillion bytes of data being produced continuously. As we might not want to store all this data and even if we could store it, it might not be computationally possible to train with this amount of data in a single go. Hence this proposal would be incomplete unless we proposed research on uncertainty estimation and robustness in the context of both continual learning and sparsification. The overarching objective for this Fellowship is to retain Prof Phil Torr within the UK and within academia. His research area of Computer Vision, and in particular deep learning, is of increasing interest to companies, as well as overseas academic institutions. The prestige and long-term funding of a Turing AI World Leader Researcher Fellowship would not only secure Prof Torr's continued commitment to UK academic research, but would also enable Oxford to build a Centre of Excellence for Robust and Trustworthy Deep Learning around him, and enable him to take a greater leadership role within Oxford, the UK and internationally.

Compilers are a crucial component of our computing stack. A compiler translates the high-level source code to low-level machine instructions to run on the underlying hardware. It is responsible for ensuring software runs efficiently so that our computers can provide more real-time information, faster services, and better user experience, and has a less environmental impact. While being a vital software infrastructure, today's compilers still rely on techniques developed several decades ago. They are limited by many sub-optimal choices used to work around the constraints of computers designed 30 years ago. As a result, today's compiler infrastructure is too old to utilise advanced algorithms and is too complex for any compiler developer to reason about successfully. Worse, existing compilers are all out-of-date and fail to capitalise on modern hardware design, causing huge performance loss and energy inefficiency. This compiler-hardware mismatch, in turn, leads to poor user experience and hinders scientific discovery and business innovation. A crisis is looming - without a solution, either hardware innovation will stall as software cannot fit, or computing performance and energy efficiency will suffer. Such a crisis requires us to rethink how we design and implement compilers fundamentally. This project aims to bring compiler technology to the 21st century to allow compilers to take advantage of machine learning (ML) and artificial intelligence (AI) techniques and modern computing hardware. Our goal is to massively reduce the human involvement in developing compiler optimisations so that compilers can quickly catch up with the ever-changing hardware to deliver scalable performance on the current and future computing hardware. We believe that ML is entirely capable of constructing efficient compiler optimisation heuristics from simple rules with zero human guidance. This idea of fully relying on ML to learn code analysis and optimisation strategies is highly speculative and has not been tested before. However, the recent breakthrough effectiveness of ML in domains like game playing, natural language processing, drug discovery, chip design, and autonomous systems gives us the confidence that this is now possible in compilers. If AI can learn to drive a car, it must be able to reason about programs to perform optimisations like scheduling machine instructions. This ambitious project, if successful, will have a transformative impact on how we design compilers. Our software prototype will be open-sourced and integrated with a key compiler infrastructure. It opens up a new way to automate the entire compiler development process, allowing compilers to get the most out of new computer hardware architecture. It will help to safeguard the massive $400B investment in today's software-hardware ecosystem and provide a pathway to greater performance in the future. The current push for specialised computer processors will not be effective if the software cannot utilise the hardware. By significantly reducing expert involvement in compiler development, this project offers a sustainable way for software to manage the hardware complexity, enabling innovation and continued growth in computing hardware. Given the accelerated and disrupted changes in hardware technology and the massive mismatch between software and hardware, success in this project will be of interest to companies that provide hardware IP and software development tools, two areas in which the UK is world-leading. It will also help ensure continued performance improvement for end-users, despite the radical changes in computer systems due to the end of Moore's Law. We believe that we have the skills, expertise, partners and work plan to achieve the ambitious goal. We are world-leading in ML-based code optimisation, have pioneered in employing deep learning for compiler optimisation and have collaborative links with key industry stakeholders in the areas.

Over the past decade, deep learning methods have had an enormous impact on the academic and industrial worlds, opening new multi-billion markets ranging from driver-less cars to speech recognition and machine translation. Deep learning has been an emerging technology for decades; it took an orchestrated scientific and engineering effort as well as harnessing of the increasing computational power and large datasets to achieve an overarching technological and societal impact. Most of the successful deep learning methods such as Deep Convolutional Neural Networks (DCNNs) are based on classical signal/image processing models that limit their applicability to data with underlying Euclidean grid-like structure, e.g., 2D/3D images or audio signals. Non-Euclidean (graph-or manifold-structured) data are becoming increasingly abundant; prominent examples include 3D objects (represented as meshes or point clouds) in CV and graphics, as well as social networks, graphs of molecules, and interactomes. Until recently, this has been a significant obstacle precluding the adoption of ML tools in some of the most promising fields. To bridge the gap between Euclidean (e.g., images, videos & speech) and non-Euclidean (e.g., graph and manifolds) ML umbrella terms have recently been coined, such as ''Geometric Deep Learning'' (GDL). Such methods have gained a keen interest in the ML community the past couple of years since graphs can model very abstract systems of relations or interactions, and thus potentially applied across the board. Recent successful examples of the application of non-Euclidean deep learning are as diverse as semantic segmentation on meshes and point clouds, drug-design and event classification in particle physics. Nevertheless, the focus is mainly on discriminative approaches (e.g., classification and segmentation problems) and limited progress has been made towards generative methodologies (i.e., unsupervised methodologies that model the distribution of data) on non-Euclidean spaces. The drawback of discriminative methodologies is that they require a massive amount of labelled, mainly manually, data, which is very expensive, or even impossible to find in many settings. On the other hand, generative approaches can operate in unsupervised scenarios and can even be used to produce data that can be utilised to train discriminative approaches. Currently, available generative frameworks have been developed primarily for Euclidean data (e.g., images, videos) and are not suitable for the non-Euclidean setting. GNoMON aims at bridging this gap by developing a mathematically principled framework for designing and implementing Generative Models for non-Euclidean domains such as graphs or manifolds. We will explore challenging problems in 3D CV and graphics. Nevertheless, the developed techniques will be designed in such a way to be general so that can aid the research in many other fields.

With the prevalence of data-hungry deep learning approaches in Artificial Intelligent (AI) as the de facto standard, now more than ever there is a need for labelled data. However, while there have been interesting recent discussions on the definition of readiness levels of data, the same type of scrutiny on annotations is still missing in general: we do not know how or when the annotations were collected or what their inherent biases are. Additionally, there are now forms of annotation beyond standard static sets of labels that call for a formalisation and redefinition of the annotation concept (e.g., rewards in reinforcement learning or directed links in causality). During this Fellowship we will design and establish the protocols for transparent annotations that empowers the data curator to report on the process, the practitioner to automatically evaluate the value of annotations and the users to provide the most informative and actionable feedback. This Fellowship will address all these through a holistic human-centric research agenda, bridging gaps in fundamental research and public engagement with AI. The Fellowship aims to lay the foundations for a two-way approach to annotations, where the paradigm is shifted from annotations simply being a resource to them becoming a means for AI systems and humans to interact. The bigger picture is that, with annotations seen as an interface between both entities, we will be in a much better position to guide the relation of trust in between learning systems and users, where users translate their preferences into the learning systems' objective functions. This approach will help produce a much needed transformation in how potentially sensitive aspects of AI become a step closer to being reliable and trustworthy.

Towards an Equitable Social VR Social Virtual Reality (SVR) constructs a digital parallel to the physical world, enabling remote social engagement mediated by modern immersive Virtual Reality (VR) technology. This social engagement is not strictly limited to conventional social interaction, but has also recently expanded to include activities such as remote participation in training, work, and service delivery. This digital parallel world offers significant opportunities for greater inclusion of individuals who are currently marginalised by the physical world, thereby widening access to the Digital Economy. SVR is a rapidly emerging technology and its pace of adoption has accelerated in the global pandemic. However, to date, there has been limited research examining the accessibility and inclusion requirements of SVR for users who currently face digital access barriers due to a disability or age-related capability loss. As a society, we sit at a critical juncture where concepts of inclusion and accessibility can be embedded into SVR while the technology is still in its formative stage. Towards an Equitable Social VR addresses the need to ensure that SVR platforms are accessible and inclusive for people with disabilities and older people, thus allowing for the potential of the platforms in contributing to the quality of life of these population groups to be realised in full. The project will undertake a programme of R&D with the aim of delivering the SVR Inclusion Framework: a collection of formalised guidance and tools serving to facilitate equal participation in SVR for disabled and older users. The project will take into account the whole spectrum of capability loss manifestations, including vision, hearing, mobility, dexterity, and neurodiversity aspects of cognition (learning difficulties) and mental health, as well as the co-occurrence of capability loss.