ABSTRACT. Being a career researcher isn’t just about “doing the research”. We’re called upon to plan what we’ll do, as well as having to pitch for - and win - support from departments, conference committees, government agencies and companies. In this talk, I’ll look at how we can make plans for our research, and in due course, how to make these plans into what we hope will be successful research grant applications. We’ll look too at some of the ways of getting it wrong, as well as tips for getting it right.

Simon Thompson is a Professor of logic and computation in the School of Computing at the University of Kent. His research has covered many aspects of functional programming, including verification, tool building and testing for Erlang, Haskell and OCaml. He is an author of books on type theory, Haskell and Erlang, and has run a MOOC on Erlang for FutureLearn.

ABSTRACT. Cloud computing is a model 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 effort. Cloud computing infrastructure includes numerous physical machines located in their data centres. Each of these physical machines runs multiple virtual machines (VMs), which are presented to the end-users. These VMs enable the cloud users to access the shared pool of computing resources. If any of these VMs gets infected by a malware, the rest of the VMs on the cloud become vulnerable to malware infection. This is because the malware-infected VM (called botVMs) can easily spread the same malware to its peer-VMs. This can lead to formation of a network of botVMs (called botcloud) which are controlled by the same attacker (called botmaster). It becomes even more difficult to detect these botVMs as they work in a distributed manner (no single-point-of-failure).

This research focuses on developing an efficient, cost-effective and scalable solution for botVM detection. The basic idea of our solution is to let VMs continuously analyse the behaviour patterns of peer-VMs for any malicious activity. By using the distributed approach of monitoring peer-VMs we reduce the risk of overloading any single VM, thus making the solution more scalable. Also, as a VM is monitored by several peer-VMs, the effect of false reporting also reduces. This improves the effectiveness of the solution. If any VM is suspected of having a malicious behaviour pattern, then an in-depth investigation is done. In this investigation traces of malicious activities called malware symptoms are searched in their memory logs, thus identifying malicious VMs. Our hypothesis is that by using this approach of two-stage detection with continuous behaviour watching facility, we may be able to detect malicious VMs more effectively and more efficiently.

This poster shows the architectural design of our botVM detection solution and its components. The next stage of work will be to investigate the effect of additional computational and communicational loads imposed on the VMs under different conditions.

ABSTRACT. Introduction: Experimental meta-data reporting is a very important field for reproducing and understanding biomedical experiments and results. Diseases caused by parasites, such as Chagas disease, are causing millions of people serious morbidity that might affect their mortality. The genetic background of the host and the parasite used in the experiments, such as the sex of the host, affects the infection outcomes. Checklist Questions (CLQs) have been designed to capture the experimental meta-data and evaluate the quality of reporting. Answering CLQs automatically is important for many reasons: CLQs allow to check completeness and clarity of experimental meta-data, and this can be used in the peer-review process. Answers to CLQs can be used to search the relevant literature for meta-data analysis process in an efficient way. However, answering the questions automatically is challenging. For example, identifying one species as the answer from many mentions of species requires an automatic understanding of the context the species are mentioned in. The research main objectives are; Explore which kind of CLQs can be answered automatically, and Combine Text Mining techniques (TM), Background Knowledge resources (BK) and the article structure to extract the answers. Method: We used 45 scientific articles from parasitology literature to answer three CLQs automatically: 1- host’s name, 2- host’s strain, 3- parasite’s name. The article title, abstract and beginning of the method sections were used to search for the answers. We utilized the CLQs keywords to search the BK and used standard TM techniques while taking into account the structure of the article. Three Named Entity Recognition tools and two large databases were used to extract the answers. Entities’ relations were extracted using a rule-based method consisting of 16 rules to match the styling of answers reporting. We calculated the co-occurrence of answers in abstract and method sections to increase the answers confidence. Answers generation method was designed to select and generate the correct answer for CLQs from various ones. Results and Discussion: 1- The current TM tools and BK resources are not sufficient alone to recognise the correct answers from the extracted entities. For instance, strains have complex nomenclature structure combining capital and small letter with numbers and punctuations. Some texts contain terms with structures similar to the strain, which decreased the accuracy of the extracted strains. 2- The host description (name, strain) is usually found together in 1-4 sentences in the beginning of the method section. 3- Both (host and parasite) or one of them were reported in the abstract and title sections. 4- The host is linked to the parasite with a verb phrase “infected with” or similar in the method section. 5- Precision (P), Recall (R), F-measures (F) and Accuracy were calculated on a question level to evaluate the performance of detecting CLQs answers, with an overall accuracy of 90%. High (P) and Accuracy show methodology confidence in detecting a checklist question's answer, and more than 50% of the detected answers were detected by rule-based relation extraction. 6- The accuracy of the correctly matched answers with gold standard data was calculated, with an overall accuracy of 80%. Conclusion: The current results show that in the current stage the methodology is limited to detect one answer for a question. We are interested to explore using all the article's sections for cases reported more than one answer. Moreover, depending on NER for relation extraction caused low accuracy when NER produces high number of false negative. We are planning for the next two years to look into two different CLQs to check the possibility of and challenges related to answering different CLQs. Also, it is an important step to evaluate the applicability of answering different questions that requires different NER and BK resources.

ABSTRACT. Depth estimation is the task of assigning a depth map to a single RGB image. Since depth sensors have become widely available, it has become possible to apply supervised learning techniques to the task and fully-convolutional neural networks currently represent the state-of-the-art in terms of accuracy.

ABSTRACT. ECG is a diagnostic and prognostic test that is widely used in clinical settings to assess heart function and detect cardiac pathologies. Some life-threatening arrhythmias, which can lead to sudden cardiac death, can be triggered by abnormal changes in the ECG including QT-interval and ST-T changes. If such abnormal changes discovered earlier, timely clinical intervention may save people's lives. With the ever-evolving wearable ECG devices, ECG monitoring became more affordable and possible to adopt outside the clinical environment more than before. People at high risk to develop these life-threatening arrhythmias, such as people who take medications that known to trigger cardiac arrest events, can benefit and utilise these technologies to self-monitoring their ECGs at home. 

Despite recent and continuous advances in automated/computerised electrocardiogram (ECG) interpretation, human interpretation remains the most accurate and reliable method. In fact, Artificial intelligence (AI) and machine learning acknowledge the significance of human supervision and involvement during each process of training and optimising AI model using active learning methods, which is recently have been known as a Human-in-the-loop approach. Besides, understanding how human perceive non-trivial patterns in ECGs would potentially contribute to the knowledge of how to train and optimise AI models in a more sophisticated way. Despite the matter of the accuracy/optimisation of detecting such warning changes in ECG, patients and caregivers need to understand the information about their cardiac health not only to self-monitor but to feel more in control of their own health and facilitate the communication with their health practitioners. 

The challenge lies in how to communicate and present this complex data to lay people, who usually do not have the training and knowledge typically associated with ECG interpretation. Data visualisation methods were developed on the assumption that they reduce the cognitive load and enhance a human's perception of patterns and trends in big and complex data. This PhD research is about understanding and modelling how people perceive abnormal changes including QT-interval and ST-T in standard ECG and how such perception could be influenced by employing data visualisations.

ABSTRACT. Echo State Networks more frequently make use of they hyperbolic tangent function as their activation function. However, this function has several undesirable properties. We explore whether there are families of piecewise linear activation functions that might serve as a better choice.

ABSTRACT. For decades, every new processor generation was characterised by significant improvements to sequential execution performance. This translated into faster execution times for any program almost automatically, with no need for a programming paradigm change. In recent years however, physical constraints led to a manufacturing shift towards multi-core processors. Whilst this brought impressive computational power, it also exposed the users to the difficulty of parallel programming. Among the possible solutions to this problem, the dataflow model of computation is, arguably, the most intuitive. First, given that the execution of instructions is triggered by the availability of their data operands, synchronization of parallel activities is implicit. Second, as the only sequencing between instructions stems from the data dependencies between them, a dataflow graph automatically exposes all forms of available parallelism. As implementing this concept at the hardware level is both inefficient and an arduous task, today's processors remain largely based on the classical von Neumann, ordered-sequence-of-instructions approach. For this reason, dataflow programs usually group instructions into larger grains, or tasks, so that instructions within them can be scheduled in the usual sequential fashion whereas only the grains themselves are scheduled in a dataflow fashion. This raises a question on the choice of tasks' granularity. Polyhedral techniques are, when applicable, invaluable for this tuning process. They first generate an algebraic representation of loop-based program parts and then work on that space to find suitable groupings, or tiles, of loop iterations. Suitable tiles will maximize data locality and, consequently, minimize data movement costs at the expense of less concurrency. Here we motivate ongoing research on how to leverage the explicit annotations already present in dataflow programs to specify inter-task dependencies, to also enable polyhedral transformations in general purpose programs of this kind.

ABSTRACT. The general idea behind sentiment analyses in finance is the existence of relevant but hard to quantify information in the textual data in addition to the objective statements of facts. This information could influence the sentiment of market participants which in turn could influence their actions. Therefore, quantifying sentiment is important to the extent that it can be informative about real future events. The traditional method used for quantifying sentiment and expectations of market participants was conducting surveys. This research aims to identify and examine sentiment indexes that are constructed from unstructured textual sources: news articles, social media messages and company event announcements.

ABSTRACT. Ontologies can present a conceptual view of a relational database. This ontology-based data access (OBDA) can allow a client to query enterprise ontologies directly. The problem of rewriting and optimisation of such queries against ontologies is insufficiently studied in database research. In this paper, we discuss using uniform interpolation to forget some symbols, especially role symbols, to rewrite instance queries against ontologies. In particular, when there is no nesting in an ontology, our forgetting algorithm is guaranteed to terminate in a polynomial time. We introduce Ackermann's lemma-based algorithm to preserve semantic equivalence during query rewriting. We further extend our approach to the guarded fragment of first-order logic. This can be regarded as generalisations of description logics, which provide bases of ontology languages.

ABSTRACT. Personal Informatics (PI) systems collect different types of personal data to support self-reflection. Individuals encounter a stage called reflection when using PI systems. This is when they try to make sense of the personal insights from the collected data which is, ideally, followed by taking actions such as behavioural change. One of the challenges that hinder reaching such goal is when people are lost in the discovery phase of reflection. Self-trackers who are in the discovery phase, typically, try to divide their high-level goals into trackable goals that PI systems can provide meaningful and actionable insights for. They also tend to collect many data because they seek to recognise which factors affect or correlate with other factors. However, most of the PI systems are well-designed for those who already know their goals but can be burdensome to those who are exploring the PI systems to identify them. Being in the latter state might create room for misinterpretations and cognitive biases and might eventually lead to “tracking fatigue”. The purpose of this research is to investigate the challenges faced by individuals who are using PI systems for exploratory purposes (i.e. in the discovery phase). Using interviews and by quantitatively analysing their eye-tracking data, we seek to identify whether individuals exhibit cognitive biases when reflecting on their data. Also, by qualitatively analysing interviews’ data, we aim to discover themes, gaps and identify implications. These artifacts should help in creating an intervention that support self-trackers in the discovery phase and improve user experience in PI systems.

ABSTRACT. In this poster, we present a tagging method that is used for generating and verifying tags of outsourced data ( i.e. integrity authenticators). Data owners generate tags of their data for integrity preservation and whenever want to check the data integrity, their tags can be verified. Comparing with existing methods, the method is secure and efficient by generating shorter and unforgeable tags. Furthermore, it can burden less computational cost of tags verification as well as less storage overhead.  the method features are achieved by using following cryptographic primitive; namely BLS short signature as well as the idea of tag deduplication.

ABSTRACT. Data fusion is a method where information from a number of datasets can be merged. In many situations, the main dataset, while required for certain purposes, may not provide sufficient amount of information in order to further process data. However, there may also exist one or more other complementary datasets, data from which can be fuse with the main dataset to provide its full explanation. This research in this project is focused on investigating the usage of data fusion of two datasets: the main numerical sensor data and industrial log files. Text files contain complementary information and are currently researched. In extracting information, there is focus on key words explaining conditions of industrial plants, events such as incidents, faults or regular maintenance and time of the events. Eventually the information will be synchronised with information in the main dataset to add signature of each parts of the dataset. 

The fusion is the main part of this project because information from two datasets must be merged in order to provide a full description of the data. There is a number of challenges. The first problem is related to the numerical data set. In order to perform supervised training, data require labels and signatures associated with fault. The current techniques and methods of anomaly detection can locate a problem. For the system to work also details of the problem must be known. Another challenge is related to the size of the numerical dataset. The data are recorded from sensors attached to plants. However, an industrial plant is a very complex device which consist of a number of elements. Each element is equipped with a number of sensors which monitor the plant performance every time interval. Because this process is continuous, the size of the data set is very large. This also causes problem when the data are analysed. The third problem related to detection of those type of faults which might not even be observed.

In order to overcome the problems, we are searching for important information in other complementary dataset, such as industrial log files. This requires research in the text mining field. From general point of view, the new data will provide significant information, e.g. signatures of a particular event which can be associated with a corresponding event in the numerical data. The locations of signals that seems to be abnormal can be explained using an additional data which provide complementary information. The text files produced by shift operators on side contain plenty of details related to incidents associated with plants

One of the research problem is to identify methods to recognise synonyms in the corpus. The words which are important are codes following a specific pattern as well as asset names related to either industrial plants or parts of the plants, e.g. pump, tank, cavity, etc. In this particular problem, code words and assets happen to be synonyms because a codeword is one kind of assets, e.g. a pump. However, the more specific lexical relationships between a codeword and an asset, such as hyponyms, could also exist if a codeword happens to be a specific kind of an asset. On the top of current challenges, the problem becomes more difficult because code words may follow a more complex pattern that reflects a part of a more complex objects, e.g. a pump in a cavity, or a pump in a tank. In terms of lexical relationships those would be meronyms and holonyms.

To fulfil this objective, methods based on Bayesian probability as well entropy are considered. Due to “noisy” entries, e.g. missing spaces, shortcuts for certain names, there are also consider methods based on text mining including spelling correction or non-symbolic analysis. One of the following objectives is recognition of periodic events related to assets followed by recognition of periodic events. It is assumed that this will lead to determination of whether events have positive or negative character e.g. whether maintenance work or fault is detected.

ABSTRACT. A learning task is sequential if the samples become available over time, e.g., computer network traffic and financial transactions. Kernel adaptive filters (KAF) are sequential learning algorithms based on kernels. In this work, we propose an automated framework for online prediction using KAF. Our proposal, unlike traditional KAF, does not need a predefined set of kernel-sizes. Here, the kernel-sizes are created in an online sequential way. The framework is tested in two real-world data sets, i.e., internet traffic and foreign exchange market. Results show that our proposal converges faster and achieve better accuracy without any specific hyperparameter tuning.

ABSTRACT. Data wrangling is the process whereby data is cleaned and integrated for analysis. Data wrangling, even with tool support, is typically a labour intensive process. One aspect of data wrangling involves carrying out format transformations on attribute values, for example so that names or phone numbers are represented consistently. Recent research has developed techniques for synthesising format transformation programs from examples of the source and target representations. This is valuable, but still requires a user to provide suitable examples, something that may be challenging in applications in which there are huge data sets or numerous data sources. In this paper we investigate the automatic discovery of examples that can be used to synthesise format transformation programs. In particular, we propose two approaches to identifying candidate data examples and validating the transformations that are synthesised from them. The approaches are evaluated empirically using data sets from open government data.

ABSTRACT. Visual Question Answering (VQA) is the task of answer any question about a given image. It has gained popularity in recent years due to its multi-domain nature: a VQA system has to be able to detect the objects in the image, parse the question and find the correlation between what has been asked and what has been seen. Current state-of-the-art methods use end-to-end statistical learning approaches to produce vector representations of the image and the question and then use them to output an answer. However, some studies have shown that this kind of methods tend not to actually capture question or image semantics, but rather rely on statistical biases learnt during training to provide an answer. The aim of this project is to build a semantic-focused VQA system based on explicit image representations.

ABSTRACT. Automatic methods for question generation from ontologies have been used in experimental contexts but they have not been used yet in practice. One of the issues that stand in the way of the generation methods being field-ready is the lack of rich ontologies and the challenging task of building ontologies by instructors who are, most of the time, not familiar with the formalism underlying ontologies or the accompanying set of tools. We focus on medical case-based question generation. We address the knowledge acquisition problem by presenting a simple text mining pipeline for information extraction from a bank of hand-written case-based questions and formalizing this information as an OWL ontology. A total of 75 questions in the specialty of pediatric were mined for axioms using the proposed pipeline. The ontology is then used to generate the same type of questions using a question generator that we developed. We will evaluate both the pipeline and the generation process through an expert review of the extracted axioms and the generated questions respectively.

ABSTRACT. In the United Kingdom, transcripts of parliamentary debates provide access to a wealth of information concerning the opinions and attitudes of politicians and their parties towards arguably the most important topics facing society, as well as potential insights into the parliamentary democratic process. However, while there exists a large quantity of research on sentiment analysis in domains such as product reviews and social media, there has been little work on applying opinion mining techniques to the domain of parliamentary speeches.

We approach this problem by examining the viability of conducting sentiment analysis in this domain at the whole speech level. We construct a new corpus of annotated debate speeches and propose a two-step sentiment analysis model. We find that, by using this model and including contextual metadata features as well as textual features, good performance can be obtained on this task, but that this approach yields limited gains in knowledge about the participants’ real opinions compared with merely inspecting their voting records.

Results suggest that a more finely-grained approach may be necessary in order to conduct a more nuanced analysis of speakers’ opinions towards the wide range of topics that they talk about in their speeches. To this end, we propose a joint sentiment-subjectivity analysis model for future work.

ABSTRACT. One of the most significant instances for making sense of healthcare data is temporal information (e.g., dates, duration and frequency). This power comes from health events (e.g., symptoms, diagnoses, and treatments) which all happen in specific time; happen in a particular order; and have a particular duration and frequency. Knowing the chronological order of events is crucial and could potentially increase the knowledge and the understanding of medical phenomena; enhance patients care; and improve medical research. Electronic health record (EHR) have facilitated the way of making these data available for clinician, researchers and decision makers. However, extraction, normalization and reasoning of temporal data in HER is difficult to achieve. This is because beside time stamped structured data (e.g., appointments and laboratory results), there is a considerable amount of temporal information in unstructured free-text format (e.g., clinical narratives, reports, discharge summaries, and treatment history). Determining temporal relations among events in free-text is a complex task due to the variation in the temporal expressions (e.g., yesterday, last night). Therefore, the main purpose of this project is to develop a methodology to integrate temporal information from a variety of healthcare data sources (structured and unstructured); and automatically infer temporal relations between events and time expressions.

ABSTRACT. Long Short-Term Memory (LSTM) networks have been demonstrated to be particularly useful for learning sequences containing longer term patterns, due to their ability to maintain long term memory. In current research we use LSTM networks for change detection in time series. A recurrent neural networks (RNNs) with LSTM units will be trained to learn the normal time series patterns and used as a predictor over one or a number of future time steps. The resulting prediction errors are modeled as a multivariate Gaussian distribution, which is used to assess the likelihood of anomalous behavior. As an initial experiment, this approach is demonstrated on time series for sin wave.

ABSTRACT. The vast amount of available data resources has made text mining indispensable. Medical documents require special information extraction techniques as information is highly correlated. Literature-based discovery is a field of text mining that can automatically extract indirect associations between concepts in textual sources. We propose a new literature-based discovery (LBD) framework, focusing on assisting cancer-related research. We illustrate our approach as a 3-stage process that can support clinicians in the identification of potential associations between concepts of interest.

ABSTRACT. This project presents an idea on constructing Aesthetical Style Learning System on Traditional Chinese Painting(TCP) and tends to develop a Style Transfer Algorithm on Generating Traditional Chinese Painting basing on the output knowledge returned by the learning system. The hypothesis of this research is that the art style of TCP could be learned and be used by the machine to generate Chinese paintings. To achieve the targets, the project has two objectives. First, to develop a TCP style learning system, which includes a TCP classifier and a TCP style learner. Second, to compute a TCP style transfer algorithm using GAN.

ABSTRACT. The availability of labelled dataset is important for supervised machine learning methods. In many cases, there are abundant unlabelled data for labelling. However, manually annotating a dataset is time-consuming and expensive. In order to reduce the human annotation workload (e.g. domain experts) and produce a labelled dataset rapidly, active learning has been proposed to minimise the cost of annotating a dataset by only annotating representative and informative instances instead of the whole unlabelled dataset. Moreover, we can also obtain labels from multiple non-expert annotators on a crowdsourcing platform in a rapid and cost-efficient manner. Our research aims to solve three problems in this scenario: 1) reduce annotation workload: how to select representative and informative instances; 2) learning whom to trust: how to know which annotators are reliable and how to aggregate labels when annotators disagree with each other; 3) training from multiple annotations: how to train a supervised machine learning model when each instance has multiple labels from different annotators.

ABSTRACT. Formal languages have been used to study computation extensively. One such language is known as type theory, the inception of which can be traced back to the work of Bertrand Russel. Recent developments in studying models of type-theoretic concepts have established deep and fruitful connections to the mathematical study of spaces, in particular the branch of topology. We explore the outlines of this nascent field.

ABSTRACT. The advances in Visual Simultaneous Localisation and Mapping (VSLAM), a key component in many vision-based and robotic systems, has been brought about by their ever-increasing algorithmic sophistication coupled with hardware computational capabilities. However, the deployment of such algorithms with their full potential into resource-constrained mobile platforms remains a challenge. This is due to the inability of these algorithms to fully utilise the available heterogeneous compute resources while meeting their performance requirements (e.g frame rate in a given power envelope) and to dynamically adjust to the variations in the perceived environment scene.

The purpose of this study is the development of an adaptive and dynamic run-time manager which aims to achieve the user-specified performance objectives for VSLAM while managing the limited resources. The study will initially involve a design space exploration (DSE) on three diverse VSLAM systems and heterogeneous resources of mobile and desktop platform. This is to identify key parameters and the significance of their variability during run-time. The run-time manager will then coordinate the execution of VSLAM on the heterogeneous resources to meet the target objectives while adjusting the run-time parameters accordingly. The poster will present initial design space exploration results.

ABSTRACT. The world is polyglot with the availability of multiple programming languages. The choice of a language is influenced by various aspects such as application's domain, programmability and level of abstraction. In the recent years, dynamically typed programming languages received wide adoption due to their ease of programming. Many such languages use virtual machine based execution approach where the program is first interpreted and the frequently executing parts of the programs are Just-in-Time (JIT) compiled. However, the performance of programs written in such languages is low compared to popular languages such as Java. This is mainly due to the level of efforts gone in development of their compilers, virtual machine (VM) or both. It is now possible for dynamic languages to reuse a production quality VM for Java using Truffle, a language implementation framework, to achieve performance as good as of Java.

On the other hand, computing hardware has become more parallel and heterogeneous. It requires programming languages to provide the ability to benefit from such hardware. To provide support for parallelism is difficult mainly due to a choice of the memory model, parallel programming approach and its portability on similar architectures. The situation worsens to support heterogeneous devices. OpenMP is addressing this challenge. OpenMP is a well-known directive based parallel programming approach for multi-core CPUs. In recent years, OpenMP has improved support for heterogeneous devices.

The poster would describe our plan to evaluate trade-offs between performance and programmability when an OpenMP based parallel programming approach is used for Truffle based language implementation. We are currently working on the Sulong project that enables the execution of LLVM IR compiled programs on an JVM using Truffle. Currently, Sulong can execute only single threaded applications and we plan to extend it to execute multi-threaded applications programmed using OpenMP. This poster will inform that, in the context of managed language runtime systems such as JVM, the incorporation of additional dynamic information will open up interesting avenues for research that may improve performance portability, scalability and load-balancing techniques for LLVM-IR compiled programs executing on multicore systems with the aid of Sulong.

ABSTRACT. Background: Mammographic density is an important risk factor for breast cancer. Recent research showed that density assessed visually using Visual Analogue Scales (VAS) showed stronger risk prediction than existing automated density measures, suggesting readers may recognise relevant image features not yet captured by automated methods.

Method: We have built CNNs to predict VAS scores from full- eld digital mammograms. The CNNs are trained using whole-image mammograms, each labelled with the average VAS score of two independent readers. They learn a mapping between mammographic appearance and VAS score so that at test time, they can predict VAS score for an unseen image. Networks were trained using > 65000 mammographic images from > 16000 women, and tested on a dataset of > 70000 images and case-control sets of contralateral mammograms of screen detected cancers and prior images of women with cancers detected subsequently, matched to controls by age, menopausal status, parity, HRT and BMI.

Results: Pearsons correlation coefficient between readers' and predicted VAS in the large dataset was 0.79 per mammogram and 0.83 per woman (averaging over all views). In the case-control sets, odds ratio of cancer in the highest vs lowest quintile of density were 3.07 (95%CI: 1.97 - 4.77) for the screen detected cancers and 3.52 (95% CI: 2.22 - 5.58) for the priors, with matched concordance indices of 0.59 (0.55 - 0.64) and 0.61 (0.58 - 0.65) respectively.

Conclusion: Our fully automated method demonstrated encouraging results which compare well with existing methods, including VAS.

ABSTRACT. Accurately representing the meaning of a piece of text, otherwise known as sen- tence modelling, is an important compo- nent in many natural language inference tasks. We survey the spectrum of these methods, which lie along two dimensions: input representation granularity and com- position model complexity. Using this framework, we reveal in our quantitative and qualitative experiments the limitations of the current state-of-the-art model in the context of sentence similarity tasks.

ABSTRACT. Spin-transfer-torque magnetoresistive random access memory (STT-MRAM) is a non-volatile technology in which data is stored as stable magnetic states of devices called magnetic tunnel junctions (device which consists of two magnetic layers separated by an insulating layer). STT-MRAM has excellent scalability and fast read/write speed, however faces challenges for future development. Challenges include reducing the switching current required to read/write data whilst maintaining thermal stability of the structure [1]. We plan to investigate the benefit of including a graded magnetic layer in the magnetic tunnel junction (MTJ) with a view to reducing the required switching current of the device [2]. To fabricate efficient MTJs, optimisation of individual layers is essential. A high quality insulating layer of magnesium oxide, MgO, is critical to achieving optimum device performance [3], therefore initial work has focussed on precisely tailoring parameters to grow thin MgO layers (2nm) using sputtering, a material deposition technique. Here, we have investigated the effects of deposition power and annealing temperature of MgO grown onto CoFeB thin films (5nm) which are representative of the magnetic layer in a MTJ. As a smooth CoFeB/MgO interface is essential [4], the surface roughness of the MgO layer was investigated. We have also characterised electron transport through the MgO using a high resolution microscopy technique, Kelvin Probe Force Microscopy. This work on the optimisation of the MgO layer will enable the later fabrication of MTJs with a graded magnetic layer.

[1] D. Apalkov et al. Proc. IEEE 104 1796-1830 (2016). [2] C. Zha et al. Appl. Phys. Lett. 97 182504 (2010). [3] S. Ikeda et al. Appl. Phys. Lett. 93 082508 (2008). [4] A. Kaidatzis et al. J. Phys.: Conf. Ser. 903 012019 (2017).

ABSTRACT. Digital neuromorphic computers are designed with the goal to simulate brain. Biophysical phenomena in the brain are mapped to differential equations which are solved on many small classical (Von Neumann architecture) processors which in turn communicate with each other through simple signals. By profiling these neuromorphic systems, we confirm that doing arithmetic operations and moving data is where the main energy and time budget is spent. But does the brain need high accuracy arithmetic processing? It is a hard question to answer because biological brain does not need to do pure arithmetic operations the same way that brain simulators do - everything in the brain is represented by physical quantities. Deep learning, currently very famous machine learning algorithm inspired by working of the brain, has been shown to be effective with limited precision when some level of stochasticity is added to round the numerical values. But does this solution generalise to Spiking-Neural-Networks simulations of various activities in the brain, far more complex than deep learning? We can build circuits and software that allows to trade arithmetic precision for latency and energy, a method usually called approximate computing. Approximate computing, both at the algorithmic level and circuit level, may enable next generation neuromorphic machines to run larger scale programmable models with more realistic biophysical neuron properties and complex brain plasticity with negligible levels of information loss due to inaccuracy in circuits and algorithms.

ABSTRACT. Increasingly, distributed computing services are deployed to process big data. MapReduce (MR), a distributed programming framework, is one of such services. It is widely adopted by public and private sectors and also served as part of other distributed computing services. It may be deployed in a private cloud, a public cloud, or a set of public clouds provided by different cloud service providers. MR may be used in a collaborative manner where datasets being analysed and users involved in the analysis are from multiple administrative domains or organisations.

Using a distributed computing service, such as the MR service, for collaborative data analysis in some security sensitive application areas, such as cyber security data analysis, could raise some serious security concerns. Tackling these concerns could be challenging if the analysis is carried out in a multiple public cloud environment as multiple public clouds are typically connected via the Internet which is open to a wide range of security threats and attacks. In addition, there are insiders in the collaborative organisations that are not fully trustworthy.

This research focuses on the security property of authentication and examines what is the most effective and efficient way of providing an authentication service to a distributed computing service in a multiple public cloud environment. To make the work feasible within the given time frame, we choose MR as the example distributed computing framework to carry out this research. This poster summarises the first piece of work that has been achieved, i.e., the design of a novel entity authentication scheme. The scheme is part of our novel Multi-domain Decentralised Authentication (MDA) framework designed as an authentication service to support the use of MR in a multiple public cloud environment. To this aim, this poster shows a generic job execution (i.e., a set of tasks to process data) model used as an execution environment for the scheme design and explains three main ideas used for satisfying the design requirements for the scheme. The next stage work is shown in the future work section.

ABSTRACT. The analysis of the network topology of complex networks with even a modest number of nodes is a high dimensional problem. Previous studies[1,2] have used principal component analysis (PCA) as a tool to spot correlations between various one-dimensional metrics of network topology, while other work[3] has investigated correlations between motif profiles on ego-networks. Our work applies PCA to higher-dimensional descriptions of network topology, such as degree-degree correlation matrices and motif profiles, across many networks to find similarities and differences between them in the hope of finding surprising global patterns. Thus far, we have tested our technique on the classical network models (Erdős–Rényi, Barabási- Albert, Watts-Strogatz), and have made some progress towards using real networks. In experiments, we can distinguish between undirected ecological networks[4] and social networks[5] by performing PCA on their 4-motif profiles. We are exploring the possibility that these or similar techniques could be used for identifying new network topology metrics. The first principal component in our experiments, for example, has large positive components for motifs containing triangles and large negative components for star-like motifs, which could be regarded as a proxy for the clustering coefficient; the technique’s ability to “rediscover” a well-known metric suggests that it could be used to find novel metrics for discriminating between network topologies.

[1] Gergana Bounova and Olivier De Weck. “Overview of metrics and their correlation patterns for multiple-metric topology analysis on heterogeneous graph ensembles”. In: Physical Review E - Statistical, Nonlinear, and Soft Matter Physics 85.1 (2012). [2] JRF Ronqui and Gonzalo Travieso. “Analyzing complex networks through correlations in centrality measurements”. In: Journal of Statistical Mechanics: Theory and Experiment (2015), p. 9. [3] Derek O’Callaghan et al. “Network Analysis of Recurring YouTube Spam Campaigns”. In: (2012), pp. 531–534. [4] RM Thompson and CR Townsend. Interaction Web DataBase. 2003. [5] Jérôme Kunegis. “KONECT: The Koblenz Network Collection”. In: Proceedings of the 22nd International Conference on World Wide Web. WWW ’13 Companion. New York, NY, USA: ACM, 2013, pp. 1343–1350.

ABSTRACT. I am developing design methodology, tools, and techniques in order to enable space-time multiplexing of reconfigurable resources which are being used in data centres and high performance computing platforms. Space-time resource multiplexing will enhance resource utilisation, therefore will improve overall system performance towards maximum throughput as well as optimal power efficiency.

ABSTRACT. The SpiNNakEar system simulates the early stages of the mammalian auditory pathway, from the outer ear to the auditory nerve. This implementation is on a neuromorphic hardware platform, the SpiNNaker machine. We exploit the massive parallelism of the target architecture to obtain real-time modelling of a biologically realistic number of human auditory nerve fibres. The results of the SpiNNaker implementation are shown to be comparable with a Matlab version of the same model algorithms, while removing the inherent performance overheads associated with an increase in auditory model scale that are seen in conventional computer simulations. Finally we outline the potential for using this system as part of a full-scale digital model of the complete human auditory pathway on the SpiNNaker platform.

ABSTRACT. Arabic calligraphy is one of the great art forms of the world. It displays Arabic phrases, commonly taken from the Holy Quran, in beautiful two-dimensional form. The use of two dimensions, and the interweaving of letters and words makes reading a far greater challenge for Artificial Intelligence (AI) than reading standard printed or hand-written Arabic. To approach this challenge, we have constructed a dataset of Arabic calligraphic letters, along with a corresponding corpus of phrases and quotes. The letters dataset contains a total of 3,467 images for 32 various categories of Arabic calligraphic-type letters. The associated text corpus contains 544 unique quoted phrases. These data were collected from various open sources on the web, and include examples from several Arabic calligraphic styles. We have also undertaken both an explorative statistical analysis of this data, and initial machine learning investigations. These analyses suggest that combining knowledge of a limited variety of Arabic calligraphy texts, with a successful machine will be sufficient for the machine reading of forms of Arabic calligraphy.

ABSTRACT. Deep learning systems can be easily fooled by small, worst-case perturbations of their inputs, known as adversarial examples. One particularly important scenario that has been overlooked is personalised medicine. To this end we introduce the concept of adversarial patients in the context of deep counterfactual models. In contrast to traditional supervised problems, in counterfactual models we can argue about the existence of adversarial examples that can fool the model in different ways. We show how small perturbations can result in different factual outcomes and we extend the concept of adversarial patients to the case of individual treatment effect estimation. We suggest that adversarial patients can serve as warning flags both during the deployment of the model but also during the design of clinical trials building a level of trust.

ABSTRACT. We present an approach to reasoning with large theories which is based on the abstraction-refinement framework. The proposed approach consists of the following approximations: the over-approximation, the under-approximation, and the combination of both.

ABSTRACT. Entity mentions embedded in longer entity mentions are referred to as nested entities. Most named entity recognition (NER) systems deal only with the flat outermost entities and ignore the inner nested ones, which fails to capture finer-grained semantic information in underlying texts. To address this issue, we propose a novel neural model to identify nested entities by dynamically stacking flat NER layers. Each flat NER layer is based on the state-of-the-art flat NER model that captures sequential context representation with bidirectional Long Short-Term Memory (LSTM) layer and feeds it to the cascaded CRF layer. Our model merges the output of the LSTM layer in the current flat NER layer to build new representation for detected entities and subsequently feeds them into the next flat NER layer. This allows our model to extract outer entities by taking full advantage of information encoded in their corresponding inner entities, in an inside-to-outside way. Our model dynamically stacks the flat NER layers until no outer entities are extracted. Extensive evaluation shows that our dynamic model outperforms state-of-the-art feature-based systems on nested NER, achieving 74.4% and 71.8% on GENIA and ACE2005 datasets, respectively, in terms of F-score.

ABSTRACT. Structural synaptic plasticity is an omnipresent mechanism in mammalian brains, involved in learning, memory, and recovery from lesions. Structural plasticity is also a useful computational tool, seeing use in automatically generating connectivity based on experimental activity data, exploring network states for Bayesian inference and assisting wider-spread synaptic plasticity rules for better performance. The structural organisation of cortical areas is not random, topographic maps are common-place in sensory processing centres. Topographic organisation allows for optimal wiring between neurons, multimodal sensory integration, and performs input dimensionality reduction. Herein we describe an efficient framework which can be used to simulate models of structural plasticity on the SpiNNaker neuromorphic system in real time in conjunction with synaptic plasticity rules, such as spike-timing dependent plasticity (STDP). A model of generic topographic map formation is implemented using our framework making use of both activity dependent and independent processes. Additionally, it can also be used to generate topographic maps between layers of neurons with no initial connectivity between them, and stabilize projections which would otherwise be unstable. Finally, we show that supervised MNIST handwritten digit classification can be performed in the absence of synaptic plasticity rules.