ABSTRACT. Currently, information, in the form of written language, is produced, shared and consumed at an extremely rapid rate. Although quantity is increasing, at the same time, the issue of quality and trustworthiness is becoming fuzzier. Whether it is scientific text or news articles, identifying the desired piece of information and assessing its credibility is becoming an involved task. Hence, it is important not only to automate information extraction but also to quantify the certainty of this information. Throughout literature there has been a variety of factors identified that influence the certainty of an extracted piece of information, ranging from expressions of speculation and hedging to timeliness, source and subjectivity.

In this case, we focus on scientific texts, to explore two main factors that can affect the credibility of extracted statement, namely, confidence expressed by the author about the truth of the statement, and trust in the author by other authors as expressed by citations. Hence we present a set of text mining methods for the extraction of textual uncertainty expressions and identification of their influence on a specific statement (event) to tackle the author confidence task. Subsequently we show how, combined with citation network analysis to estimate trust in the author(s) we can provide a more accurate measure of uncertainty for a given statement. Finally, we demonstrate efficiency and performance for scientific applications.

ABSTRACT. Dataspaces aim to remove the up-front costs of information integration by gathering the needed domain information through targeted interactions with the end-user throughout the life-time of the integration. State-of-the-art tools are used to rapidly construct an initial (incorrect) integration, which is then refined in a pay-as-you-go manner by asking end-users to supply feedback on the resulting data. The idea is that end-users will choose to put effort into providing feedback on the areas of the integration where the quality is important to them, while other less well-used areas will receive a smaller share of user attention. This approach is promising but open problems remain. One issue is that the end-user loses control over the process. Their contribution is to specify their query requirements and to provide feedback on the results, as directed by the dataspace. But what feedback should the user supply to get the data they want? We propose a new approach to data integration in which the end-user and the dataspace work as equal partners to meet the integration goal. Both are able to perform data integration tasks directly, and both request and provide feedback on the results. In addition, the dataspace observes the actions of the end-user when carrying out integration, with the aim of automating that part of the work in future integration tasks. In this paper, we explore this idea by examining how a dataspace can observe an end-user at work, correcting errors in query results, to gather feedback needed to refine the mappings used for integration. We propose an algorithm for converting manual corrections to feedback and present the results of a preliminary evaluation comparing this approach with seeking explicit feedback from end-users.

ABSTRACT. There is an increasing demand on Emergency Department (EDs) services, in England alone there has been 22.3 million attendances recorded in 2014-2015, with an increase of 2.7% from the previous year and an increase of about 25% over the last decade. 30-50% of A&E attendances are referred for radiographic examination. These X-rays are initially interpreted by ED medical staff, however the accuracy of these interpretations has always been a source of concern. This is mainly because the majority of patients attending EDs are seen by junior doctors. This problem is widely acknowledged, so in many hospitals X-rays are reviewed by an expert radiologist at a later date - however this can lead to significant delays impacting on the eventual outcome.

Focusing on the most common diagnostic errors in EDs: "missing wrist fractures" I started my PhD journey on the field of medical image understanding and ,in this talk, I will share some of my interesting findings and insights.

ABSTRACT. Speech emotion detection and recognition technology (SER) plays an important role in achieving the advanced-level intelligent agent. A bottleneck in SER is that in a cross-corpus setting, where the training and testing data are very different due to the influence of language, emotion eliciting way (acted or natural), speaker difference, etc., a pre-trained recognition system may degrade greatly in the testing stage. To address this issue, we treat this difference as domain shift and propose using adversarial models to align the distributions of different emotion corpora. Specifically, we assume one source corpus/domain, which has plenty of labeled data, and one target domain, which we are really interested in but suffers from a lack of data and a shift from the source domain. The connection between the two domains allows for the knowledge transfer, and particularly, adversarial models are employed to bridge the shift so that the target-domain recognition performance should be improved.

ABSTRACT. Osteoarthritis(OA) is the leading cause of disability worldwide. Despite its prevalence and copious effort by the research community OA is still a poorly understood condition. Some evidence on risk factors has been found but limited treatments are available. This project aimed at developing machine learning techniques to automatically assess the severity of this disease just by looking at plain radiographic images of the knee. Being able to detect this disease fast and cheaply can lead to the implementation of therapies and change in lifestyle, which improve mobility and potentially slow down disease development and its symptoms. Furthermore, we looked at future developments of the disease determining to what extent it was possible to determine if healthy people were likely to develop OA in the future. Finally, we investigated what imaging features are most related to different kinds of pain measurements and looked at more consistent ways of assessing it.

ABSTRACT. It is crucial that modern industrial plants maintain a safe environment in which processes can run efficiently. To achieve this, processes are being monitored via sensors and associated alarms. When an abnormality in the system is detected an alarm is raised and transmitted to a control room to be interpreted by a human operator. But given the rapid increase in the number of alarms used on each process this task has become very difficult, especially in case of alarm floods. Thus the major goal of current research in alarm system monitoring is to aid human operators in identifying and answering to faults detected in the industrial plant. Initially much emphasis was put on improving the physical aspects of alarms, but lately, because of the overwhelming amount of data, a necessity for numerical analysis methods has become clear. Most focus has been on using discriminative methods to identify and classify floods. In this talk, however, I would like to bring your attention towards a model based approach to alarm data analysis by building a dynamic Bayesian network to capture conditional activation dependencies between alarms. To show the potential of this model, besides flood identification and classification, examples of root cause detection and data generation will be given. The latter being of special importance as it solves the issues with data availability in the field caused by the proprietary nature of most real data and the lack faulty data.

ABSTRACT. Rapid technological advances in genotyping and sequencing technologies are driving the generation of vast amounts of genomic data. This presents a unique opportunity to improve our understanding of the environmental and genetic mechanisms that give rise to phenotypes. This data is technically hard to analyse because there many attributes (often in the order of a million), and a large amount of background knowledge is relevant. Genotype data are most commonly used in genomic models to identify genetic regions which control phenotypes, and to predict the likelihood that members of a population will produce progeny with particular phenotypes. However, most of the data may be irrelevant for certain phenotypes, leading to suboptimal, difficult to understand models. To meet this challenge, we propose a three-stage general framework that incorporates background knowledge in its model building processes by applying feature stability, inductive logic programming, and meta learning. We demonstrate that the proposed framework (1) significantly outperforms the state-of-the-art in identifying associated genomic regions, (2) identifies relevant genetic configurations, and (3) improves overall phenotype prediction, using a diverse Rice (Oryza sativa) population.

ABSTRACT. Reaching Exascale poses tremendous challenges to the computer architecture community. In ExaNeSt we are developing a novel computing architecture designed to meet these challenges across the entire hardware and software stack. New technologies are being leveraged to produce a small-scale, 2-cabinet prototype. We are currently designing an advanced computing rack, consisting of densely-packed FPGA MPSoCs with in-node distributed NVM storage and immersion liquid cooling. Here we present an overview of our FPGA-based custom-made top-of-rack switch that envisions a paradigm shift by avoiding the costly and inefficient routing tables used in most current HPC systems. These tables are the biggest consumers of chip resources (i.e. area and power) in the switching elements and, indeed, they are usually much larger than the router logic (i.e. buffering, flow control mechanisms plus the crossbar) as they need thousands of entries to sustain large number of endpoints without harming the performance. Instead of routing tables it uses an arithmetic routing policy based on the positions of the router in the Network. Our switch design re-purposes a portion of the saved routing table area to integrate Virtual Output Queues, which improve significantly the performance of the network by reducing the likelihood of congestion being formed. The switch is low-latency, taking only 3 clock cycles (with no contention) and uses a custom-made packet format devised within ExaNeSt to provide a design which will scale to over a million nodes.

ABSTRACT. There is a huge demand for on-device execution of deep learning algorithms on mobile and embedded platforms. These devices present constraints on the application due to limited resources and power. Hence, developing energy-efficient solutions to address this issue will require innovation in algorithmic design, software and hardware. Such innovation requires benchmarking and characterization of Deep Neural Networks based on performance and energy-consumption alongside accuracy. However, current benchmarks studies in existing deep learning frameworks (for example, Caffe, Tensorflow, Torch and others) are based on performance of these applications on high-end CPUs and GPUs. In this work, we introduce a benchmarking framework called "SyNERGY" to measure the energy and time of 11 representative Deep Convolutional Neural Networks on embedded platforms such as NVidia Jetson TX1. We integrate ARM’s Streamline Performance Analyser with standard deep learning frameworks such as Caffe and CuDNNv5, to study the execution behaviour of current deep learning models at a fine-grained level (or specific layers) on image processing tasks. In addition, we build an initial multi-variable linear regression model to predict energy consumption of unseen neural network models based on the number of SIMD instructions executed and main memory accesses of the CPU cores of the TX1 with a test error rate of 8.04 ± 5.96%. Surprisingly, we find that it is possible to refine the model to predict the number of SIMD instructions and main memory accesses solely from the application’s Multiply-Accumulate (MAC) counts, hence, eliminating the need for actual measurements. Our predicted results demonstrate 7.08 ± 6.0 % average relative error over actual energy measurements of all 11 networks tested, except MobileNet. By including MobileNet the average relative test error increases to 17.33 ± 12.2 %

ABSTRACT. Because of the widespread use of the internet, many companies have acquired an online presence which allows them to gather large amounts of information about the people that use their services. This results in the creation of large, hard to maintain, and difficult to access databases. These systems have traditionally been managed in software, but in recent years the use of hardware to accelerate databases has become a viable solution. There are already several methods of accelerating databases using hardware (Intel- ligent storage controllers, GPU, ASIC), but we are proposing the use of FPGAs (Field Programmable Gate Arrays) for the task. FPGAs are mostly used as ASIC (Applica- tion Specific Integrated Circuit) substitutes, but that does not make full use of their potential, which lays in their ability to adopt the functionality and structure of different digital circuits at run-time. We propose using the FPGA’s partial run-time reconfiguration capabilities in order to get more use out of the on chip logic and allow the chip to perform a larger amount of tasks in database applications. We will see that this is possible and beneficial when used together with regular software databases to accelerate queries. In particular, we will focus on an interface between the software database and the FPGA accelerator that facilitates the use of hardware to speed up execution.

ABSTRACT. ARM released 64-bit support for their processors in a new Instruction Set Architecture called A64 in a separate execution state (AArch64).

Dynamic Binary Modification (DBM) is a technique that allows an application to be modified at runtime transparently, as it executes at the machine instruction level. DBMs are used in many areas such as virtualisation, instrumentation, program analysis and translation. However, DBM tools generally suffer from high overheads due to the additional complexity of modifying the native code dynamically.

In this work, we present a low overhead and optimized DBM tool for 64-bit ARM processors. Our results show our DBM tool achieves less than 15% overhead compared to native execution on the SPEC CPU 2006 (a single-threaded benchmark suite) and even lower on PARSEC 3.0 (a multi-threaded benchmark suite).

ABSTRACT. In Behaviour-Driven Development (BDD), the behaviour of the software to be built is specified as a set of example interactions with the system using a "Given-When-Then" structure. These examples are expressed in high level domain-specific terms, and are executable, acting thus both as specifications of the requirements as well as tests that can verify whether the current system implementation matches desired behaviour or not. This approach has many advantages but also poses some problems. When the specifications grow large, they can become costly to maintain and extend. Specifications can grow larger than necessary, because of the difficulty of spotting duplication in collections of hundreds of examples, developed by a continuously changing team over a period of time. We propose to use dynamic tracing to identify duplication candidates in BDD specifications. Specifically, we focus on the patterns of production code methods invoked by each example, and consider two examples to be duplicates if they invoke the same functionality in the system under test. Empirical evaluation of our tool on three open source software systems shows promising results.

ABSTRACT. Abstract. Data Wrangling (DW) is the subject of growing interest given its potential to improve data quality. DW applies interactive and iterative data profiling, cleaning, transformation, integration and visualization operations to improve the quality of data. Several domain independent DW tools have been developed to tackle data quality issues across domains. Using generic data wrangling tools requires a time-consuming and costly DW process often involving advanced IT knowledge beyond the skills set of traffic analysts. In this paper, we propose a conceptual approach to data wrangling for traffic data by creating a domain-specific language for specifying traffic data wrangling tasks and an abstract set of wrangling operators that serve as the target conceptual construct for mapping domain-specific wrangling tasks. The conceptual approach discussed in this paper is tool-independent and platform agnostic and can be mapped into specific implementations of DW functions available in existing scripting languages and tools such as R, Python, Trifacta. Our aim is to enable a typical traffic analyst without expert Data Science knowledge to be able to perform basic DW tasks relevant to his domain.

ABSTRACT. Deep inference is a feature of a class of proof systems similar to sequent calculus but with the property that inference rules can be applied 'deep' inside of formulas. While proof systems which exhibit the property of deep inference are studied for variety of reasons the focus of this project is the relationship between deep inference syntax and other mathematical structures, in particular category theory. It has become clear that various features of a deep inference system can be viewed as common categorical structures. The focus of my research is on deep inference systems for linear logic which for certain fragments already has nice categorical interpretations and well understood notions of proof-equivalence. This allows me to draw on the existing work in linear logic to study to study categorical properties of deep inference systems.