ABSTRACT. SpiNNaker is a multi-core digital simulation platform designed for running real-time simulations of Spiking Neural Networks (SNNs).  A human brain processes information from the outside world in the form of 'spikes' - but before any of this can happen the environmental stimuli must be converted into a spiking signal by specialised sensory organs. The cochlea is the sensory organ used in hearing, it converts sound into spikes using many sensing cells all working in parallel. This work uses a biologically faithful model of the human ear mapped onto a SpiNNaker machine (SpiNNak-Ear) and demonstrates interaction with subsequent auditory pathway nuclei in a complete simulation.

ABSTRACT. SpiNNaker is a 1 Million ARM processor machine for simulating neural networks. As floating-point support is expensive in circuit area and energy, integer processors were chosen and fixed-point arithmetic is utilized in order to simulate real numbers. Fixed-point types cannot at the same time represent very large and very small numbers and it is always a trade-off based on specific equations and quantities that are being simulated. In this talk I will present various new techniques to improve the accuracy of fixed-point solvers for neural model ODEs and also draw some conclusions how this influences the next generation SpiNNaker2 hardware accelerators and software.

ABSTRACT. The Sulong project facilitates execution of single-threaded C programs on Java Virtual Machine (JVM) using Truffle-GraalVM ecosystem by converting them to LLVM Intermediate Representation (IR) [1]. We are extending the Sulong project to execute parallel programs written in C using OpenMP. OpenMP is a popular directive based parallel programming approach for C, C++ and FORTRAN programs. This involves implementing OpenMP runtime in Java that results in mapping OpenMP memory model on top of the Java memory model. As of now, we have implemented an initial version of the runtime that can execute benchmarks from NPB suite correctly [2]. Currently, we are optimising the implementation for its performance. In the future, the support for heterogeneous devices using Tornado framework could be added [3]. Additionally, we are working on performance analysis of Truffle hosted guest languages such as Sulong for LLVM IR, GraalJS for JavaScript, TruffleRuby for Ruby to identify where the time is spent. Existing Java profilers cannot distinguish between the execution of a Java program and a guest language program. To demonstrate the viability of the approach, we have adapted the existing profiler, perf-map-agent, to map guest language function information to their respective Java methods. It helps to map profile for JVM-based execution to the execution of a guest language program. We plan to extend this work to analyse the performance of OpenMP programs on JVM by recording thread-level execution using perf.

[1] https://github.com/oracle/graal/tree/master/sulong [2] https://www.nas.nasa.gov/publications/npb.html [3] https://github.com/beehive-lab/Tornado

ABSTRACT. FPGA accelerators are being deployed at a large-scale for their performance and energy efficiency for many applications. However, currently, the ecosystem does not provide enough support for FPGAs to be utilized to their full capacity. One problem is that resource allocation is often constrained to a run-to-completion model or acceleration of a single application at a time. In this talk, I will describe how we can virtualize FPGA accelerators in Space-Time domain using resource elasticity to allow dynamic allocation of reconfigurable resources for higher utilization and performance, transparently from the user. I will also briefly mention how we can use the same techniques to enable live migration as well as heterogeneous scheduling for CPU+FPGA platforms.

ABSTRACT. This talk presents the motivation behind the project, work out difficulties and requirements to deploy reconfigurable resources on high performance computing (HPC) systems and data centres. Dynamic partial reconfiguration for exascale computing allows to move FPGA acceleration to data and provides the baseline to resilience by allowing runtime adaptations of the system. Moreover, the here presented partial reconfiguration method allows integrating modules more flexibly that what is able with the FPGA vendor tools and can ease the overall design of acceleration system by decoupling the system design from the core accelerator modules. With this, we are providing an accelerator-centric view to application designers. This talk will also look beyond HPC to address problems related to FPGA virtualisation as needed for recent large scale FPGA cloud deployments.

ABSTRACT. SpiNNaker is the world's largest neuromorphic computing platform capable of simulating up to 200 million neurons in real time. We use SpiNNaker to explore a wide-spread learning mechanism present in mammalian brains: structural plasticity. Neuron constantly rewire their connectivity in response to changes in input statistics. While the process is very hard to explore experimentally, we make it easily available as a learning rule in the sPyNNaker software for running simulation on SpiNNaker.

ABSTRACT. Question Answering (QA) is defined as retrieving the correct answer given a question in natural language. With the resurgence of deep learning techniques and their application to natural language processing (NLP), some QA systems are capable of surpassing human performance. Recent analyses show, however, that these systems are likely to exploit syntactic patterns inherent to the associated evaluation challenges. This talks presents efforts towards alleviating those problems and introduces an approach for QA over unstructured text that builds upon a novel family of learning algorithms which learn over data in relational format (Graphs). Relational information is injected into unstructured text in form of background knowledge, syntactic analysis and linguistic theories.

ABSTRACT. Data Integrity Auditing (DIA) is a security service for ensuring the integrity of data stored in a third-party based storage service using integrity tags. We propose a novel tagging method, called Tagging of Outsourced Data (TOD), for the generation and verification of tags. The design of TOD makes a hybrid use of multiple cryptographic primitives, namely homomorphic encryption, algebraic signature and BLS short signature, as well as the ideas of tag deduplication and decoupling block indices from tag generations. TOD has a number of unique properties: (i) it supports both public and private verifiability, (ii) it protects data confidentiality and is resilient against tag forgery, (iii) it supports dynamic data operations. Performance evaluation indicates that the method imposes less storage and computational overheads than related methods.

ABSTRACT. Monitoring and understanding behavioural aspects from smartphone data provide an opportunity to improve people's health and well-being. However, developing systems that detect behavioural aspects to generate personalised health feedback is an ongoing research area. In this project, we seek to recognise the motivating interests from smartphone's data and use that to nudge people by providing personalised recommendations to change unhealthy behaviour.

ABSTRACT. Case-based multiple choice questions MCQs are a popular form of assessment in the medical domain. Having the capabilities to process human-authored questions automatically and understand their content opens up new possibilities in generating variants of existing questions, extracting templates used to generate new types of questions, forming theories about the difficulty of questions, or the other way around, evaluating difficulty models on existing questions. It also support question designers in classifying and retrieving questions based on their characteristics. Finally, it also helps in analysing different aspects of their quality such as incompleteness or ambiguity in their feedback.

Case-based MCQs are well structured and involve a large number of semantic relations. They are also characterised by heavy use of anaphoric expressions and sentences that express relations with one relation argument being implicit (i.e. can be inferred using structural clues). In this paper, we focus on relation extraction in diagnostic case-based MCQs. We analyse the structure of these questions, highlight issues of coreference and the absence of explicit mention of one relation argument and propose a rule-based approach to overcome these issues. Key features of our approach are the use of question structure to resolve coreference and the sentence classification component that is used to infer the implicit relation argument.

The evaluation of the approach has focused on the correctness of extracted relations. Correctness was assessed through expert judgement and finding support for the relation in existing ontologies.

ABSTRACT. Parkinsons Disease (PD) is an incurable long-term neurological disease with an unknown cause. It affects one percent of the population of people over sixty. PD patients usually spent more time reside in their homes and have less social interactions. So social withdrawal is one of the potential symptoms caused by PD. Social wellbeing has been shown to be essential to the quality of life of PD patients. But little is known about how the progression of PD influence social life. And the concept of social withdrawal is not clear from various literature. There is also no widely used social withdrawal measurement, especially for PD. Smartphones are now widely used and with advanced sensors embedded, they can detect heterogeneous data of behaviour of the user. This project will utilize smartphones as a monitoring tool for measuring social withdrawal. Having working definitions and practical mapping method for social withdrawal is necessary. This presentation discussed how we extract the practical interpretation of it and build a feasible assessment from the disease and the smartphone perspective.

ABSTRACT. There are numerous toxins and medications that can cause ECG changes, even in patients without history of cardiac pathology. Certain medications can produce a complication known as `drug-induced long QT syndrome', shown on the ECG as an elongated QT-interval. Self-monitoring for this could be lifesaving, as the syndrome can increase the risk of a life-threatening arrhythmia, known as torsades de pointes (TdP), that can lead to sudden death in young, otherwise healthy people. However, ECG interpretation is known to be complex, even for clinicians, and as such little work has examined self-monitoring. This research aims to examine lay people's ability to identify a drug-induced QT-interval prolongation on an ECG, and to determine whether using data visualisation techniques can support them to self-monitor.

ABSTRACT. With the use of manual annotations and statistical analysis techniques, scientists have demonstrated that a considerable portion of mobile app reviews are relevant to app evolution. Based on such demonstrations, variety of automatic classification and extraction approaches are proposed to assist developers in parsing and processing user reviews. However, there is a major concern among both scientific community and development teams over the quality and usefulness of user reviews from developers’ point of view. Justifying that a review is relevant to software evolution does not necessarily indicate that it is useful for developers. Therefore, it is not incorrect to argue that, despite attempts of existing exploratory studies, the usefulness of user reviews from developers’ viewpoint is still obscure. Accordingly, effectiveness of existing extraction tools developed based on such hypothesis is under question. In this project, useful mobile app reviews from developers’ perspective are elicited by definition of a set of criteria and application of it using qualitative content analysis techniques. An automatic model is then to be proposed to assess the usefulness of a mobile app review based on the predefined criteria. Performance of the model would be examined on real world data and fine-tuned accordingly.

ABSTRACT. Abductive reasoning generates explanatory hypotheses for new observations using prior knowledge. We present a forgetting-based approach to ABox (ground) abduction in description logic (DL) ontologies. The hypotheses computed by the method make the fewest assumptions necessary while taking the form of a disjunction of ABox axioms, where each disjunct is an independent explanation for the observation. We explore the elimination of redundancies from hypotheses in practice, resulting in an approximate and an exhaustive approach, both of which are evaluated using a prototype implementation over a corpus of ontologies. We also discuss current directions of research, such as the use of other forgetting methods, hypothesis refinement and the integration of abduction and learning in DL ontologies.

ABSTRACT. Our interest is ontology-based query answering for the Horn loosely guarded fragment. Given a Boolean conjunctive query (BCQ) q, a database D and a set of Horn loosely guarded formulas F, our target is to check whether D and F entails q. The decision procedure is based on ordered resolution with selection, and has been proved to be sound and refutationally complete. The result shows that this approach can decide BCQ answering for the Horn loosely guarded fragment.

ABSTRACT. Reasoning higher-order logic has historically lagged behind that in first-order logic. This is due to its inherent complexity. However, numerous domains could benefit from powerful native higher-order reasoning. Higher-order logic is the natural language for much of mathematics and facilitates reasoning about functions and programs. This presentation explores treating higher-order logic as a first-order theory and how this can be done efficiently.

ABSTRACT. The early phases of requirements engineering (RE) deal with a vast amount of software requirements (i.e.,requirements that define characteristics of software systems), which are typically expressed in natural language. Analysing such unstructured requirements, usually obtained from stakeholders’ inputs, is considered a challenging task due to the inherent ambiguity and inconsistency of natural language. To support such a task, methods based on natural language processing (NLP) can be employed. In this talk, we describe a new resource, i.e., embedding-based representations of semantic frames in FrameNet, which was developed to support the detection of relations between software requirements. Our embeddings, which encapsulate contextual information at the semantic frame level, were trained on a large corpus of requirements. Compared with existing resources, our embeddings provide encouraging results to describe, and hence, organise software requirements within RE domain.

ABSTRACT. Data stream management systems exist to support dynamic anal- ysis of streaming data, often to inform decision-making. Decision support systems exist to enable decisions to be made that take into account user priorities. However, although these categories of system are now quite mature, there has been little work in- vestigating their use together. We bring together a well established streaming platform (Storm) and a widely used decision-support methodology (Analytic Hierarchy Process) to provide dynamic decision support over data streams. In so do- ing, we also investigate approaches making recommendations auditable (using provenance) and trustable (using explanations). The resulting stream decision support system is illustrated using an application that supports train journey planning.

ABSTRACT. Routing protocol design is very important in network design, because an effective routing protocol can greatly improve network performance. As heavy and aggregated load and interference will decrease the performance, to balance load and interference, an effective load- and interference-balance hybrid routing protocol (LIB-HRP) is proposed. Unlike existing hybrid routing protocols, LIB-HRP considers the load condition at interfering nodes, because the interfering node with heavy load will have much influence on the current node. The heavier load condition at interfering node can bring a longer duration of interference. Besides, as mesh clients have limited energy, energy condition is taken into account for mesh clients in the proposed approach. Simulation results show that LIB-HRP can obtain better performance in terms of average packet loss ratio, delay, network throughput and energy consumption.

ABSTRACT. Classification is a core reasoning task of Description Logic (DL) Reasoners. For traversal-based reasoners, classifying the ontology involves n^2 numbers of expensive (worst case N2ExpTime) subsumption tests. Various optimizations of traversal algorithms have been designed and proved to be highly effective in reducing the number of subsumption tests (e.g. Enhanced Traversal (ET) algorithm). In this paper, we present a novel such optimization, based on the Atomic Decomposition (AD). We empirically evaluate its effectiveness with some interesting results.

ABSTRACT. Differential Evolution (DE) is a common metaheuristic that is robust to noise and capable of optimising a time-varying objective function. Despite this, it has rarely (if ever) been applied to hand tracking. In this talk, I will discuss the potential advantages and disadvantages of using DE as a hand tracking algorithm.

ABSTRACT. Adverse Drug Event (ADE) is a harmful consequence of medication errors such as drug overdose and allergic reactions. Therefore, ADE detection are crucial to prevent harmful effects of drug therapy. This talk presents a simple but efficient LSTM-based model for identifying ADEs, related drugs and their attributes such as form and dosage from discharge summaries.