ABSTRACT. My research is based on visually guided-behavioral animal modeling by using spiking neural networks to realize practical robotics.

ABSTRACT. Understanding how humans formulate their decision to use a certain technology is a key issue in the heart of Information Systems research. Two prominent models, i.e.  Technology Acceptance Model (TAM) and Unified Theory of Acceptance and Use of Technology (UTAUT), have presented extensive research addressing this phenomenon. However, the development of Machine Learning (ML) algorithms brings to the literature a more objective approach enabling the predictability of end-user’s decision of technology use. Unlike the traditional technique widely adopted in TAM and UTAUT research, the application of ML modelling accurately predicts the decision of end-users. In this research Decision Tree Regression, Linear Regression, K-Nearest Neighbour Regression, Support Vector Regressor, and Multi-Layer Perceptron are utilised to model the technology use. Formulation of the proposed model involved extracting new features from Twitter utilising Twitter API as an unobtrusive data measure enabling the objective evaluation of end-user’s perspective of technology use. The resulting model is evaluated on two major steps: features reliability using factor and discriminant analyses and modelling performance on the basis of RMSE, R-Square and MAPE metrics. The sought after goal is to develop a better model than TAM and UTUAT in terms of predictability and parsimoniousness.

ABSTRACT. The discovery of synthetic antiferromagnets (SAF) has triggered a major research effort leading to the discovery of the GMR and the birth of spintronics. Basic SAF structures are composed of two magnetic layers separated by a non-magnetic spacer layer (Fig. 1 (a)). By tuning the thickness of non-magnetic layer, the interlayer exchange coupling between the two magnetic layers can be changed from ferromagnetic (parallel) to antiferromagnetic (antiparallel) [1]. The tunability of SAF due to spacer layer thickness has huge potential in spintronic sensor and memory devices [2]. Recent publications on the potential of SAF to support skyrmions present a new direction for skyrmion-based devices. Due to the skyrmion Hall effect (SkHE), skyrmions in magnetic multilayers move with a transverse velocity [3] (Fig. 1 (b)). It has been shown that in a SAF structure, the creation of one skyrmion in the top layer will lead to a skyrmion nucleation in the bottom layer due to AFM coupling. AFM-coupled skyrmions are capable of suppressing SkHE protecting them from annihilation by touching the edge of the nanotrack [4]. In this work, we aim to develop and understand the characteristics of skyrmion-based devices on such synthetic structures. The SAF stacks will be fabricated by magnetron sputtering and characterised via X-Ray Reflectivity (XRR) /X-Ray Diffraction (XRD), Atomic Force Microscope (AFM) and Scanning Electron Microscope (SEM). The technique we use to image skyrmions is Scanning Transmission X-ray Microscope (STXM) which provides the capability to investigate magnetic properties with high spatial resolution (~10 nm) and has been previously used to demonstrate nanoscale room temperature skyrmions [5]. This work will offer a new method of controlling skyrmion and utilising different stack structures for skyrmion-based devices.

ABSTRACT. Gene expression is the key to understand genetic diseases, such as cancer. However, Many factors regulate gene expression in the cell such as tissue type, epigenetics and signals from the surrounding environment. Moreover, it can be affected by the activities of other genes which form a complex dynamic network of interaction with other molecules in the cell. This complexity must be taken into account when designing drugs against cancer; a drug that targets a specific gene by promoting or inhibiting it would usually yield a cascade of reactions inside the cell which results in a new state reflected in gene-expression levels variation. In our research, we aim to build meta-models to predict gene expression levels after exposing cells, which is taken from different types of cancerous cell lineages to various drugs in diverse conditions.

ABSTRACT. Skrymions are non-trivial particle-like magnetic structures and can be generated through varies external excitations, including gradient magnetic fields, spin-polarised currents and even localised laser irradiations [1], due to the underlying competition between short- and long-range energy interactions. By virtue of the antisymmetric Dzyaloshinskii-Moriya interaction (DMi) [2], skyrmions exhibit unique and robust topological configurations, being lied on the intersection between magnetism and topology, and therefore research in this area has become an increasingly hot topic. The desired non-volatility of skyrmions increases their potential as a candidate for future information storage devices. Using numerical simulations, we have found the radical annihilation occurs in B20 FeGe skyrmion lattice (SKL) driven by sweeping magnetic field. Due to their topological properties, unwinding of skyrmionic textures always accompanied by nucleation of Bloch points and propagation through the whole layer. During this process, the positive DMi dominates “skyrmion tube” starts to detach from bottom surface, as well as the leaking of topological charge. Moreover, different annihilation modes can also be found under the effects of the thickness-dependent demagnetisation field, Gilbert damping parameters and external applied field. On the other hand, by artificially inducing nano-engineered defects such as notch structure at the top surface, the localised asymmetric demagnetisation field significantly decreases the Bloch-point nucleation energy barrier, and as the result, the “skyrmion tube” starts to generate from the top surface. Moreover, during the annihilation process, the magnitude of generated emergent electric field is analysed, which can reach up to the order of 0.1 MV/m, and voltage between two surfaces is up to the order of 0.1 mV with frequency about 4 THz. The high stable tera-hertz frequency can also provide insight information during the skyrmion annihilation, and also delivers ideas for potential detection methods in the future experiments. These findings provide a basis insight for spintronics’ dynamic phenomena and relevant topology behaviour, which can contribute to the stability studies on skyrmionic textures and industrial developing research and development on skyrmion-based spintronic devices in the future.

ABSTRACT. Abstract— Consumers and users of websites have demonstrated little interest in reading privacy policies. Policies are usually explained in several pages and are mostly irrelevant to user’s concerns. Moreover, most privacy policy notices are written beyond the reading capacity of adults. Although Internet users are concerned about their privacy and would like to be informed about the privacy controls they can exercise, they are not willing or able to find these choices in policy text. This paper reports on an ongoing research about automatically assessing compliance of website privacy policies with GDPR to alleviate users’ concerns and bridge the above-mentioned gap. The research has sofar analysed the existing related work and has identified the gap. At the current step, we are designing an appropriate NLP based approach to sove the formulated problem.

ABSTRACT. Videos with a high level of compression often show compression artefacts such as blockiness and corrupted blocks in frames (due to errors in transmission). This poster describes a method of improving the quality of the video using image registration and the combination of frames. It uses the fact that points in the scene appear in different pixels in different frames (reliant on each frame being taken from a slightly different viewpoint).

ABSTRACT. Magnetic skyrmions (Fig. 1(a), (b)) are topologically-protected spin textures, which exhibit particle-like properties. They were recently realized in both bulk non-centrosymmetric magnetic materials and thin multilayer films with the Dzyaloshinskii-Moriya interaction (DMI) generated by strong spin-orbit coupling (SOC) and broken inversion symmetry. Novel computational and storage devices using skyrmions or skyrmion gases present with potential for improvement over traditional CMOS technology due to their robustness, nanoscale size and non-volatility. Investigation of such skyrmion systems would not only yield insight on fundamental physics, but also impact on future technologies. Magnetic skyrmions have recently been demonstrated experimentally at room temperature for the first time in technologically relevant multilayers which opens the way for their use in novel nanocomputing applications. Nowadays, magnetic skyrmions have been proposed for developing Boolean gates and full logic systems, utilising the flow through nanowire tracks and leveraging the rich physics of skyrmions: the spin-Hall effect, the skyrmion-Hall effect, skyrmion-skyrmion repulsion, skyrmions-boundaries repulsion, and electrical current-control of notch depinning. In addition to traditional logic computing and storage technologies, skyrmion-based devices exhibit the potential for multi-state functionality due to their particle-like nature. Such multi-state systems are paramount in realising neuromorphic computing technologies including artificial synapses, artificial neurons, spiking neural networks (SNN) as well as reservoir computing. Inspired by recent work on skyrmion-based neurons and synapses, we investigate here computationally the conditions for synaptic activity based on trains of skyrmions (Fig. 1(c), (d)). This work aims to develop novel low-power non-von Neumann computing paradigms utilising nanoscale skyrmions.

ABSTRACT. An investigation into the high statistical significance of Small-Angle X-ray Scattering (SAXS) in comparison to Scanning Electron Microscopy (SEM) was performed on ordered magnetic nanostructures. Such highly ordered nanoarrays play a key role in many aspects of modern science and technology, such as in artificial spin ices, as resonant enhancers in plasmonics and in magnetic data storage drives as Bit Patterned Media (BPM). The fabrication process involves several steps, which inevitably lead to a certain distribution of structural parameters. It is therefore essential to know parameters including mean diameter, standard deviation of the diameter and the positional misalignment. These parameters can then be used to optimise the fabrication process or in theoretical calculations and simulation to predict the performance of such arrays, for example in BPM. Ordinarily a real space imaging technique such as Atomic Force Microscopy (AFM) or SEM is used for analysing nanostructures. However, these serial methods are limited in the number of islands that can be measured in a single image, resulting in appreciable measurement times. Consequently, the statistical significance of these experiments is often difficult to obtain, especially in applications with large numbers of nanoislands (10^5). We report on the use of SAXS to investigate a very large number of nanoislands simultaneously, reducing measurement times to the order of seconds. This provides a method of characterising large arrays of highly ordered nearly uniform nanoislands with high statistical significance. A systematic series of samples was measured with the principal variable set as the periodicity of the islands, varying between 50 nm and 250 nm. The individual island diameter ranged between 15 nm and 40 nm with the largest islands corresponding to the arrays of greatest period. This resulted in arrays with diameter/pitch ratios varying between 0.16 and 0.30, which covers a significant part of the range of ratios encountered in the literature.

ABSTRACT. The atomic interfaces between ferromagnetically ordered thin films with perpendicular anisotropy and films with either in-plane anisotropy or a nonmagnetic metal are responsible for a range of exciting physics leading to potential devices in areas as diverse as broadband THz generation [1], DMi and Skyrmions [2], Spin Hall effects [3], and data storage [4]. Understanding the magnetic structure of these atomically engineered interfaces as a function of depth is a key element in optimizing layer structures and developing the underlying physical insight necessary for exploitation. Here we present polarised neutron reflectometry (PNR) data obtained from exchange spring structures comprising a Co/Pd superlattice with perpendicular magnetic anisotropy (PMA) and an in-plane anisotropy NiFe layer. This study builds on our prior work to characterize [Co/Pd]5/Pd/NiFe exchange spring structures using vector magnetometry, in which an orange-peel coupling was found to provide a significant contribution to the interlayer exchange for small (<2 nm) Pd spacer layer thicknesses [5]. To lend insight to the evolution of the sample magnetisation as a function of depth, PNR data is fitted to dynamical, spin-resolved simulations of the structural and magnetic scattering length densities (SLDs). Measurements made at room temperature and 200 °C are designed to determine the temperature dependence of the exchange interaction. This is an important consideration in any system which experiences an increase in temperature, either by design such as in data storage media for HAMR or as a side effect (i.e. through Joule heating) of applying a dc or ac excitation current.

[1] T. Seifert et al. Nat. Photonics 10, 483 (2016) [2] A. Fert et al. Nature Rev. Materials 2, 17031 (2017) [3] J.B.S. Mendes et al. Phys. Rev. B 89, 140406 (2014) [4] C.W. Barton et al. Sci. Rpts. 7, 44397 (2017) [5] C. Morrison et al. J. App. Phys., 117, 17B526 (2015)

ABSTRACT. In this work, we plan to investigate the possibility of integrating theory reasoning in the Inst-Gen calculus in iProver. Doing so would enable us to harness the power of dedicated theory solvers to potentially make iProver more efficient when running certain kinds of problems. Inst-Gen is notable for employing modular reasoning which allows the use of fast “off-the-shelf” solvers for propositional reasoning, therefore such an integration is a natural extension of the method.

ABSTRACT. Finding supporting evidence for addressing complex mathematical problems is a difficult task. The formalization of mathematical problems and the emergence of automated reasoning methods have brought a significant advance to this area, but the process of formalizing proofs and theorems still requires great human effort. In this work, we use mathematical theorems and proofs written in informal (natural) language in order to create a word embedding for mathematical discourse and evaluate it using the premise selection task, recommending those premises that most likely will be useful to prove a particular statement. This is a novel approach to the problem of premise selection and to the best of our knowledge, is the first to combine natural language mathematical text and deep learning for the task of premise selection. This work also contributes by introducing a novel dataset, NL-ProofWiki, to support the evaluation of premise selection systems using natural language corpora. The proposed method uses an RNN-based embedding model defined over the combination of co-reference graphs and local linguistic/equational patterns within the proofs.

ABSTRACT. Named entity recognition (NER) is a natural language processing (NLP) technique for identifying entities present in a text. To solve semantic ambiguity in natural language requirement (NLR), we decided to recognise NLR entities and analyze them in their context of use. However, there are no NLR domain specific datasets. We developed a gazetteer of entity categories and a corpora trained on deep learning algorithm for accurate tagging and entity recognition.

ABSTRACT. Silicon Photonics is considered a key enabling technology for scaling High-Performance Computing systems into the exa-scale domain. Large-scale optical switches are key components for delivering scalable optical interconnects for High-Performance Computing. However, scaling using Silicon Photonics is inhibited by significant challenges in terms of optical losses and complexity. In this work, we examine the scalability potential of an optical, thermally/electrically tuned switch based on Mach-Zender Interferometers. We propose three hardware-inspired routing strategies that leverage the inherent asymmetry present in the switch components in order to reduce optical losses and switch energy consumption. We evaluate the model using 8 realistic workload types for an increasing system size, from 16 up to 1024 endpoints. The routing strategies offer an optical loss reduction of up to 38% with respect to the absolute maximum in the best case, while the strategy with the most impact offers an 18% reduction in bit-switching energy consumption compared to the baseline in the best case.

ABSTRACT. The metamagnetic phase transition of FeRh from antiferromagnetic (AF) to ferromagnetic (FM) is a template for first order magnetic transitions [1]. The spin-flip of the anti-parallel Fe atoms induces a moment on the Rh atoms accompanied by an isotropic lattice expansion of ≈ 1%. The transition can be driven via excitation by magnetic, electronic, or thermal means (with a transition temperature (TC) ≈ 360K [2]). The intrinsic speed limit for the creation of ferromagnetic order on small length scales (100nm) was found to be ≈ 0.5ns from time and spatially resolved X-PEEM measurements [2]. It has been proposed in the literature that spin canting of the Fe and Rh moments upon excitation may be the key drivers in the transition [3], and our ultimate goal is to the determine the role of the lattice expansion in the transition. The effect of the latent interfacial FM phase present in thin films [4] depends on film thickness. The presence of anisotropic strain fields (due to lattice mismatch between MgO and FeRh) alters the magnetic order in these strained regions. These regions exist over a wide range of temperatures, including those below TC where the bulk material would be wholly AF. This project will use a laser pulse to induce the transition while the film is monitored by magneto-optical microscopy. The samples were characterised by XRD and VSM. A series of film thicknesses (10-50nm) will be investigated over a range of temperatures (80-500K) to build on previous studies [2], [5]. The range of film thicknesses and temperatures will allow us to investigate how strain affects the propagation of FM order upon laser excitation and the subsequent relaxation to the AF phase (≈ 2ns as observed by Ünal [2]). These considerations are important in applications such as heat-assisted magnetoresistance (HAMR) or AF anisotropic magnetoresistance (AFM-AMR).

ABSTRACT. In October 2017, W.H.O reported that there are 253 million people estimated to live with vision impairment; 36 million are totally-blind, and the remaining have moderate to severe low vision. This enthuses researchers to devise tools, and applications to assist visually impaired and blind people (VIB) to explore and navigate their surroundings. Navigation systems can be classified into (1) Macro-navigation, which includes: positioning, path-planning, environment representation and interaction with the user; and (2) Micro- navigation, which helps the VIB to avoid obstacles in the immediate physical environment. The most common assistive tools used by VIB are the white- cane and guide-dog, which have a lot of limitations. Thus, the need to develop a system that can provide assistance to a broader layer of the VIB where balancing between affordability and efficiency is considered.

ABSTRACT. The main aim of this project is to develop a phone-based application to assist blind people with safe navigation. The application will seek to determine the accurate location of users through the use of a phone camera, in addition to using computer vision techniques. To locate a user’s position, a building photographed by the user’s camera will first be recognized by a matching algorithm. The algorithm will attempt to match the captured photograph with a corresponding photograph stored in a database. After matching two images, the user’s location is calculated on the basis of the transformation between the two images.

ABSTRACT. it is noticeable in dierent heterogeneity types that complexity is inherent in heterogeneous data and several data analysis methods are used to work on a single data type or only a few types simultaneously. There are two main approaches to handle mixed-type datasets. The rst approach is unifying data types for all the variables (such as the continuous numerical data) and then apply the machine learning analysis. However, this approach degrades the data quality as some original data types are converted to other types in the learning stage; The second approach is to analyse a subsets that have a specic data type. Restricting the analysis process to predened data types mean that the analysis process is performed for a subset of information. As a consequence, only one part of object information is analysed, and essential relationships between features are ignored. This study considers data heterogeneity with the various data types associated with the features describing information. In contrast to the existing approaches of transferring all features to the same data type or selecting the features with the same data type, the proposed study will develop the machine learning methods (in particular feature selection and learning of regression models) based on the original heterogeneous data.

ABSTRACT. In our age of big data and the knowledge society, effective processing and sharing of knowledge is crucial. Agents provide a key abstraction within modern development of AI systems and software. A multi-agent system is a collection of semi-autonomous processes known as ‘agents which live in dy- namic and unpredictable environments. Agents continuously interact which each other and the environment, and have the ability to communicate with other agents and users. In current agent platforms a simplifying assumption is that agents speak the same language, which is recognized as a limitation. In this project the aim will be to study aspects of agent communication and knowledge sharing where this assumption is relaxed. In particular, our aim is to develop a new agent communication platform and support for sce- narios where multiple agents are responsible for different knowledge bases (e.g., that capture the agents different expertise) and have the ability to re- strict and adapt their knowledge so that it can be conveyed and understood by their communication partners. The project will study and combine methods from artificial intelligence, agent-based systems, ontology-based information systems, distributed knowl- edge, logic and automated reasoning, and is expected to make use of relevant tools such as the OWL API and LETHE that have been developed in the School.”

ABSTRACT. A novel framework for evolving spiking neural networks by composing graphs using a cooperatively evolved population of sub-graphs

ABSTRACT. Parameter tuning and optimisation of SNN models have applications in both biological models and in the conversion of artificial neural networks (ANNs) models to SNN models. Experimental limits to the data that can be collected in vivo mean that SNN models of biological neural networks often require several parameters to be estimated or tuned. The SNN models that result from the conversion of ANNs could be optimised and so help to develop better conversion methods. An evolutionary algorithm (EA) optimisation framework for SNN models on SpiNNaker was developed with a view to understanding how the weight parameters of a small test network could be optimised for the MNIST digit recognition task. A genetic algorithm (GA) was used in the experiments for its biological relevance and the potential for model evaluation to be parallelised by running multiple models simultaneously on SpiNNaker. Tools such as those developed in this research could help bridge between the fields of machine learning and computational neuroscience and allow for a better understanding of the mechanisms of information processing in neural networks more generally.

ABSTRACT. This work builds a Spiking Neural Network(SNN) on SpiNNaker for the number recognition task. Compared with other published works, this SNN uses LIF model with relatively large leaky while still gets loss-less results. Besides, the number of synapses is pruned by 55%, which makes it suitable for portable devices to achieve power efficiency.  

ABSTRACT. Our reliance on the correct functioning of embedded systems is growing rapidly. Such systems are used in a wide range of applications such as airbag control systems, mobile phones, and high-end television sets. As Internet of Things (IoT) is now present in all technology sections, allowing different embedded systems to connect and exchange data, the chances of security breach have expanded to a large extent. Thus, the reliability of the embedded (distributed) software in IoT is a key issue in the system development. This Poster is concerned with identifying software vulnerabilities, which compromise not only network security but also information security in IoT devices. In particular, it aims to find software vulnerabilities using fuzzing and symbolic execution techniques, in order to prevent unauthorized access to the network by shielding the network from malicious attacks and thus protecting the data flowing through the network.

ABSTRACT. A Mobile Ad hoc Network (MANET) is a network of wireless mobile devices capable of communicating with one another without any reliance on a fixed infrastructure. A Mobile Medium Network is a set of mobile forwarding nodes functioning as relays for facilitating communication between the users of this Mobile Medium. The performance of the Mobile Medium depends on the Mobile Medium node density, distribution and movement. In the proposed dynamic node movement, the movement is determined based on whether the node is on a forwarding path for a data flow or not. Simulation results show that slowing down the speed of mobile nodes when they are forwarding significantly affects the delivery rates in Mobile Medium networks. For networks with a few forwarding nodes dispersed in a large region reducing the mode movement speed by 50% results in an approximately 20% improvement in the delivery ratio, with even higher improvements possible at lower speeds.

ABSTRACT. ADVANCED PERSISTENT THREATS (APT) are one of the most sophisticated cyber-attacks that have been growing in recent years. APT are developed skilfully by well-resourced attackers which makes the task of detecting them extremely challenging due to their capabilities, complexity, and the deployment of obfuscation techniques (e.g. Encryption, Zero-day Vulnerabilities, and Social Engineering). Ranging from the reconnaissance to the data exfiltration, APT follow organised multiple stages to achieve their aims. This PhD aims to improve the detection of this type of cyber-attacks by providing a framework that can analyse network traffic and changes in user behaviour by using unsupervised machine learning in a real-time approach. Subsequently, abnormal behaviour extracted from the framework are exchanged between hosts in networks. In doing so, the framework helps to identify compromised devices and stop the spread of attacks on other devices in the network. Also, it minimises the manual investigation and improves the incident response by showing more details about abnormal behaviour to network administrators.

ABSTRACT. Our aim is to enable automated, quantitative analysis of pigment composition and application techniques without recourse to the actual pigment samples, but by sampling from the data itself. The documents are from the John Rylands Library Special Collection. The University of Manchester Library holds one of the finest collections of rare books, manuscripts, and archives in the world. They span five millennia and are written on virtually every medium ever employed.

ABSTRACT. To provide long-term companionship robots need to adapt to their users. Personality is a very important trait of human beings, dictating how we interact with the environment and the people around us. It is of vital importance for a robot to understand the personality of its users, to adapt to it and make the interaction with them more pleasant and engaging. The question that we are investigating with this work is: can a robot learn how to distinguish personality traits from real-time interactions with humans?

ABSTRACT. As we witness the benefits of participating in MOOCs, we address the problem of knowing if users are actually learning. The traditional approaches involve administrating tests, quizzes, self-assessment surveys, etc. We explore if is there a way to measure knowledge acquisition and personalise the users’ learning experiences in an unobtrusive manner. My Ph.D. proposes data-driven methods to measure learning by mining user interaction data to identify regularities that could be indicators of knowledge acquisition.

ABSTRACT. Reinforcement learning (RL) is constrained in its applicability to domains with virtually unlimited numbers of interactions available due to the sample inefficiency of many RL methods. An approach to remedying this issue is through the transferring of knowledge across tasks; this is often described as multi-task learning. In this poster I present an case study into the benefits and drawbacks of using low-level compared to high-level knowledge representations for multi-task reinforcement learning.

ABSTRACT. Social Internet of Things (SIoTs) is an emerging advanced paradigm of IoT, combining social networking and IoT. In the paradigm, relationships between different IoT devices are established to efficiently cooperate each other to achieve a common goal without intervention of human beings. There relationships could be co-location, co-ownership, co-type and parental-hood. The social relationships between IoT devices are research concerns in the fields. However, there lacks in aspects of human beings in SIoTs.

ABSTRACT. Deep reinforcement learning has proven to be a great success in allowing agents to learn complex tasks. However, its application to actual robots can be prohibitively expensive. Furthermore, the unpredictability of human behavior in human-robot interaction tasks can hinder convergence to a good policy. In this paper, we present an architecture that allows agents to learn models of stochastic environments and use them to accelerate learning. We descirbe how an environment model can be learned online and used to generate synthetic transitions, as well as how an agent can leverage these synthetic data to accelerate learning. We validate our approach using an experiment in which a robotic arm has to complete a task composed of a series of actions based on human gestures. Results show that our approach leads to significantly faster learning, requiring much less interaction with the environment. Furthermore, we demonstrate how learned models can be used by a robot to produce optimal plans in real world applications.

ABSTRACT. Magnetic skyrmions are topologically-stable vortex-like spin textures which form in magnetic systems that lack inversion symmetry [1, 2]. In 2016, it was demonstrated that skyrmions can be nucleated in magnetic multilayers at room temperature [3,4], giving promise to the notion of incorporating these topologically-stable spin textures into devices which can perform abacus-like calculations. Initial experiments required the use of synchrotron-based X-ray microscopy techniques in order to study both the static and dynamic properties of skyrmions. However, recent advances now mean that detection and manipulation of skyrmions in laboratory conditions is possible via electrical measurements and Scanning Probe Microscopy (SPM) [5, 6]. The current focus of the project has been to develop methods for the high-quality patterning of magnetic multilayers using argon ion milling and electron-beam lithography. Samples in the literature prepared using such techniques have shown minimal magnetic degradation, a key aspect in fabricating skyrmion-based devices [7].

[1] S. S. P. Parkin, et al., Science, 320 (5873), 190-194, (2008) [2] A. Fert, et al., Nature Nanotechnology, 8, 152–156, (2013) [3] C. Moreau-Luchaire, et al., Nature Nanotechnology, 11, 444–448, (2016) [4] O. Boulle, et al., Nature Nanotechnology, 11, 449–454, (2016) [5] W. Legrand, et al., Nano Lett., 17 (4), 2703-2712, (2017) [6] D. Maccariello, et al., Nature Nanotechnology, 13, 233–237, (2018) [7] K. Zeissler, et al., Nature Nanotechnology 13, 1161-1166, (2018)

ABSTRACT. Network protocol security requirements become more significant. However, finding effective approaches for testing network protocol security has proven to be a difficult problem. Fuzzing is one important network protocol security test methods. However, fuzzing is generally unable to creating various inputs that exercise all paths in the program. On the other hand, Symbolic execution is also one of the best technique for testing and finding vulnerabilities in the network protocols, but symbolic execution may also not achieve high path coverage because of dependence on SMT solvers and the path-explosion problem. Therefore, the purpose of this research is to combine these techniques, so we can cover many paths. In other words, we can raise the chances to find the vulnerability in the network protocols.

ABSTRACT. A study of using MPI and OpenCL programming languages to utilise a cluster of 8 FPGAs.

ABSTRACT. The Internet has become essential and the most important thing in people's life. Almost two billions people use the Internet to communicate with each other whether sending or receiving emails, using social media applications, playing games and so many more. Since the use of the Internet is increasing rapidly, another area is appearing to use the Internet as an international platform for allowing objects and smart devices to communicate, compute and process information. This platform is called the Internet of things (IoT). IoT is a massive number of things or smart objects which communicate with minimum human interventions. Nowadays, the need for smart devices is dramatically increased. These smart devices became essential in our daily lives. However, the smart devices are objects which can be connected with each other through Wifi, Bluetooth and radio frequency identification (RFID). These devices have embedded sensors or RFID tags to send and receive data. IoT covers different areas and applications such as smart home, cities and smart health care. According to Singh et al. The number of objects in IoT excluding computers, tablets, and cellphones will create $300 billion revenue until 2020. Singh et al. also said by 2020 the number of cellphones and tablets will be around 7.3 billion devices. These objects will establish a massive and complicated network with a huge amount of information being communicated over this network. Most of these smart objects connected to the Wireless network and not equipped with effective security measures and vulnerable to different privacy and security issues such as confidentiality, integrity, authenticity and access control. One of the ways to achieve data confidentiality, integrity is by using data encryption. Due to the limitations in IoT devices such as power consumption, low memory life, less processing speed, the encryption techniques have different proprieties than normal encryption. Thus, Researchers are creating different types of lightweight encryption techniques to be adequate to IoT devices by modifying key size, a number of rounds and so on.

There are many problems of lightweight encryption techniques, one of them is to verify the software. In other word, developers want to check whether the software has bugs or deep errors or not. One of the common techniques for detecting bugs and deep errors is Fuzzing and symbolic execution. Fuzzing is an automated software testing technique which involves providing invalid or unexpected as inputs. Symbolic execution is a technique for detecting vulnerabilities and generating tests. Symbolic execution uses input values instead of concrete values which is the major concept of this technique. It is assumed by the end of this research, fuzzing and symbolic execution will be the most effective technique for detecting vulnerabilities in the lightweight encryption algorithm.

ABSTRACT. The internet of things (IoT) has a rapid growth over the past decades and affect human life in different domains. Wearable devices have recently become very popular because of their unique capabilities to closely monitor an individual's activities and collect users' vital signs in real-time such as heart rate and body temperature.

Despite of the useful features of the wearable devices, users have very little (if not zero) control over the collected data, which negatively affects their privacy. The data collected are usually personal and sensitive, hence unathorised access to such data could pose serious risks to users.

The protection of individuals’ privacy is an increasing concern. Therefore, it is important to provide solutions that help users understand how their data is being used and give them control over how their data is being shared.

ABSTRACT. Research has fashions, just like everything else. In this paper, we present a trend detection algorithm and apply it to 2.7 million computer science abstracts gathered from DBLP. The resulting trends, which span thirty years, appear to have a characteristic shape when plotted on a popularity-time graph: a steep upslope, a plateau, and then a slow tail-off as the term begins to feel less “novel” to researchers and funding bodies.

ABSTRACT. Policy preferences are the positions that Members of Parliament (MPs) hold towards different issues and policy areas. They consist of information about the topics discussed in debates (such as education, law and order or defence), as well as policies towards those topics (like increasing or decreasing expenditure, or strengthening or weakening laws).

Successful identification of these preferences is a key step in analysis of the sentiments, opinions and positions expressed in Parliament, and could allow citizens to more easily monitor the work of their elected representatives.

Using natural language processing and machine learning methods, I explore supervised and unsupervised methods of topic and sentiment analysis of parliamentary debates, and present analysis of MPs' expressed policy preferences.

ABSTRACT. With heart disease remaining a leading cause of death around the world, it's important that we all monitor and maintain good cardiovascular health. We usually rely on the knowledge of nurses and doctors to inform us of our heart health, but what if we could take control of it ourselves? The Hearty app for smartphones provides a platform for recording blood pressure readings, determining and tracking heart health and providing a reminder of what medication to take and when. The Hearty app uses the well established QRISK score for calculating how healthy a user's heart is from various metrics about themselves, such as age, weight, cholesterol and blood pressure reading. The app lets the user track their blood pressure and QRISK score over time, allowing them to see the fluctuation with weight or if they have taken their medications. The app will remind the user at specified times to take their medication.

ABSTRACT. Good heuristics are essential for successful proof search in automated theorem proving. Finding such heuristics is a challenging task as the parameter space of theorem provers is vast, and diverse problems require different heuristics. This task can be split into two: finding efficient heuristics and selecting relevant heuristics based on problem properties.

ABSTRACT. Ontologies are generally understood as sets of axioms. Performing ontology engineering tasks on an unstructured collection of axioms is often cumbersome and error-prone. Such difficulties may be mitigated by an extensible meta-language based on macros for Description Logics. A framework for using macros may allow for an efficient, flexible, and adaptable way of working with ontologies in practise. 

ABSTRACT. During undergraduate courses, a student may often seek help from websites: searching for answers to coursework questions. Such interaction will change the normal learning process for students. This research will investigate student interaction with websites during undergraduate programming courses, and whether such interaction alters the student learning process.

ABSTRACT. The push for lower cost, smaller size, and more features have activated combing analog circuits with digital ones. This kind of mixed-signal chip is typically composed of sensitive analog/RF circuits and high switching frequency digital circuits. The noise generated by the switching activities can be easily coupled into the analog part through the power grid or substrate. In fact, if some encryption implementation such as Advanced Encryption Standard (AES) is being executed, the sensitive key information could be coupled into the RF circuits, amplified and transferred by transmitting circuits. Attackers could recovery the key through collecting electromagnetic(EM) traces emitted at the transmitting end. Here it is the first time to verify the feasibility of EM attacks through substrate against an AES hardware implementation. To evaluate the possibility of AES encryption key leakage through substrate noise coupling, state-of-the-art techniques to model, simulate and analyze noise coupling would be applied. 

ABSTRACT. The advancements in the Internet of Things (IoT) and Artificial Intelligence (AI) have led to the rising popularity of SH. SH offers life-enhancing services such as home automation, security and safety, utilities, healthcare, and entertainment. To provide the smart services, a range of heterogeneous devices are integrated into the home environment. These devices generate and consume, and may store an enormous amount of data including sensitive data, e.g. health information. They may also perform sensitive operations such as disabling an alarm system. If a malicious or curious user gains unauthorised access to such devices, he/she may manipulate the devices and/or steal sensitive information, compromising the privacy of the home residents or putting their safety at risk. Therefore, controlling the access to home devices to ensure that the devices can only be accessed by authorised users is paramount. However, existing access control (AC) solutions have not properly considered the SH characteristics, making them less efficient when being applied to an SH environment. This research attempts to address this open issue by designing an AC model that can securely and efficiently manage who can access which devices under what conditions.

ABSTRACT. With advances in technology, desktop virtual environments (VEs) have become a powerful tool for interactive visualization and immersion of the user in 3D virtual space. However, for these tools to be effective, they must be easy to use and a major component of usability, particularly important in desktop VEs, is navigation, a task that users often find difficult; despite having many controllers to navigate VEs, not all combinations of input and display devices can work together in a suitable and useful manner as they physically need complimentary setup, for instance.

The aim of this research is to present a new multimodal interaction mapping framework for 3D object manipulation within the virtual reality (VR) realm, by leveraging the advantages of having multiple DoF (Degrees of Freedom). In this new software engineering designed framework, interaction devices such as the keyboard, mouse, joystick, and specialist devices for 3D interactions; the Wing and the 3D connexion space navigator, can all be combined to provide a more intuitive and natural command system. This can be applied to many different specific systems including industrial applications within the petroleum, geology and materials sciences.

ABSTRACT. The aim of the research is to investigate how Knowledge Graphs (KGs) can be used to augment user cognition to address knowledge-intensive analytical and learning tasks, it will focus on the design and evaluation of efficient VR mechanisms to support the interaction with large-scale Argumentation Graphs (AGs). So that the research project allows Virtual reality technology to facilitate interaction with a 3D-AG which is based on Walton’s argumentation schemes and then allow users to interact, make a query and filter the nodes using virtual reality VR and natural language query NLQ technologies.

ABSTRACT. Antiferromagnetic (AF) spintronic devices are an active area of study at present, and FeRh is one of the most promising materials for these devices. There is a need for stable, non-volatile devices to provide room temperature storage, and metamagnetic phase change materials present a possible solution in the form of antiferromagnetic anisotropic magnetoresistance (AF-AMR) memory[1]. The AF state produces no stray magnetic fields and is itself insensitive to externally applied fields, however in the ferromagnetic (FM) state the electron spins can be aligned by an external field due to the presence of a net magnetic moment in the lattice[1]. Materials such as FeRh are able to be reversibly and reliably transitioned between the AF and FM states using controlled temperature changes[2]. This project aims to characterise FeRh devices by identifying local transition temperatures and comparing them with the topography of the device to determine any correlation. Both thin films and nanowires have been considered, with the wires expected to present increased asymmetry over films[3]. FeRh is an antiferromagnet at room temperature and transitions to a ferromagnet at around 370K, with the transition temperature approximately 10K higher during heating than cooling[4]. The phase does not change uniformly, with the FM state forming nucleations which expand to cover the whole device as it is heated[5]. The characteristics of the nucleation process are highly repeatable on subsequent heating cycles, implying that the local transition temperature is affected by the sample structure and topography in some manner. The transition has a secondary effect that the electrical resistivity of the sample is significantly different between the two phases[4], allowing the phase to be measured via electrical measurements. The local resistivity of a 60nm thick FeRh film was mapped out using TUNA while the sample was heated to a range of temperatures.

[1] X. Marti et al., Nat. Mater. 13, 367, (2014) [2] M. Fallot and R. Hocart, Rev. Sci. 77, 498, (1939) [3] J. Arregi et al., J. Phys. D: Appl. Phys. 51, 105001, (2018) [4] J. Kouvel and C. Hartelius, J. Appl. Phys. 33, 1343, (1962) [5] C. Baldasseroni et al., Appl. Phys. Lett. 100, 262401, (2012)

ABSTRACT. The present research aims to investigate whether stereotypical views of devices (Mobiles and PCs) impacts the type of information best recalled from them. Previous research has found that opinions of what a device is typically used for can impact decision making. However, this effect has not been investigated in relation to other cognitive domains. The present poster proposes a methodology to investigate whether this effect exists in regards to long-term memory recall.

ABSTRACT. This project proposes a novel reinforcement learning (RL) algorithm which is guided by a knowledge-based user feedback. The project integrates user feedback using structured knowledge bases into the traditional RL models. The main objective of the project is to reduce the barriers for a continuous, knowledge-driven dialogue between end-users and RL models, improving few-shot learning capabilities (the ability to generalize from fewer examples).

The core research questions are:

• Can a Knowledge Base driven reinforcement learning (RL) approach supported by end-user feedback deliver few-shot learning capabilities?
• Which Knowledge Representation models and formalisms can support generalization and user feedback process?
• How semantic parsing methods can be used to support the end-user interaction with the knowledge base?

The research can impact the ability to deliver Artificial Intelligence (AI) systems which are more transparent, and which generalize from fewer examples. These two properties are at the center of the requirements for the application of AI within scenarios which require trust (e.g. health and legal domains).

ABSTRACT. Question answering (QA) defines the task of finding the correct answer for a question from a given knowledge source. The current state of the art approaches employs deep learning frameworks. But these approaches are data hungry and struggles to generalise beyond the provided data. Meanwhile, cognitive and pedagogy literature suggests that explanations are a fundamental aspect of learning, helping humans better generalise the knowledge they have learned. We aim to investigate the task of incorporating explanations to improve generalisation and inference capabilities on question answering.

ABSTRACT. Question Answering (QA) is the task of finding correct answers for natural language questions in a given knowledge source. Current Deep Learning methods for QA demonstrate several limits when the task requires performing complex multi-hop reasoning and
providing explanations for the output. Applications in contexts with high social and economic impact (e.g. health care, legal settings, scientific discovery) require the evolution of current "black-box" predictors in the direction of robust reasoners able to justify their answers to the end user.We investigate the task of generating explanations for science questions, exploring the application of novel neuro-symbolic techniques for reasoning over hybrid semantic knowledge representation.

ABSTRACT. In this work, we analyse how Artificial Intelligence (AI) and Distributed Ledger infrastructure scan provide a fundamental change in the structure and dynamics of the organisations. The works of R.H. Coase and M. Olson, on the nature of the firm and the logic of collective action, respectively, are revisited under the light of these emerging new digital foundations. We also analyse how these technologies can affect the fundamental assumptions on the role of organisations (either private or public) as mechanisms for the coordination of labour. We propose that these technologies can fundamentally affect: (i) the distribution of the rewards within an organisation and (ii) the structure of the transaction costs. These changes bring the potential for addressing some of the private-public sector trade-offs.

ABSTRACT. Although it was introduced in the 70s, Symbolic Execution has found renewed interest in recent years. It has only recently been made practical, as a result of significant advances in constraint satisfiability and of more scalable, dynamic approaches, which combine concrete and symbolic execution. However, the impact of this technique is still limited. For instance, when a program executes a trace symbolically, it simulates all possible paths, which results in a large number of paths that need to be analysed “path explosion”, which caused a huge number of queries to the constraint solver that are often large and complex when analysing real-world programs. Investigating these challenges will show promise to impact symbolic execution technique from a bug finding perspective. At this research, we will investigate whether a method that combines all of fuzzing, symbolic execution and abstract interpretation techniques in a novel manner is able to detect software vulnerabilities in Unmanned Aerial Vehicles. In addition, to investigate how to mitigate their weakness such as scalability and deployment challenges that have prevented wide adoption in practice.

ABSTRACT. The Internet of Things (IoT) is a dynamic environment in which things can communicate and interact with each other. It is poised to be the new era of communication, perhaps the new Internet. In spite of the high level of convenience delivered by the IoT, it is accompanied by a range of security challenges. One of the challenges is how to achieve identification and authentication among a large number of heterogeneous devices in an automatic, secure and efficient manner. This project is set to investigate how to achieve authentication in IoT environments.

ABSTRACT. The rapid proliferation of insecure IoT devices gave rise to the IoT botnets. These botnets were used to perform unprecedented distributed denial-of-service (DDoS) attacks. This study focuses on finding an effective and efficient mechanism for tackling IoT botnets. The IoT botnets can be tackled in two ways: 1) by identifying vulnerable IoT devices and protecting them from malware attacks; and 2) by detecting infected IoT devices and isolating them.

ABSTRACT. We aim to develop an efficient global ensemble model to improve the prediction performance of heterogeneous\distributed models from distributed data sources to solve the centralised data mining issues and individual classifier limitations with high prediction accuracy. Also, to ensure the accuracy and diversity of the combined prediction models with less computational complexity, low communication costs, and storage requirements. The results of the experiment will show the effects of model selection and combining strategies on prediction accuracy and help to identify the importance of model evaluation method before combining it to the ensemble model.

ABSTRACT. Estimation of treatment effects is common in areas such as economics, marketing and healthcare. We focus on the task of performing feature selection in this setting, a task that has been connected with various desirable properties such as decrease in computational complexity and better interpretability. In particular, we focus on information theoretic approaches and show how they can be derived under a clearly defined objective and a causal graph that describes the relationships between pre-treatment features, treatment and the corresponding potential outcomes. We propose estimators useful for identifying relevant features for the partially observed potential outcomes under common treatment assignment scenarios. We then extend the theory to identifying predictive features -- i.e. indicators of heterogeneous treatment effects which can guide the design of tailored treatment assignments.

ABSTRACT. Human cells are incredibly complex, tens of thousands of genes, proteins, and small molecules interact in complex temporal-spatial ways. The only way to disentangle and model these interactions, and hence treat cancer, is through hypothesis formation, and reproducible experimentation. However, progress is slow because of the limited number of qualified human scientists. A Robot Scientist is a physically implemented laboratory automation system that exploits techniques from the field of artificial intelligence (AI) to automatically execute cycles of scientific experimentation. The Robot Scientist Eve will be used to work with human cancer cells, the aim is to test statements taken from the scientific literature.

ABSTRACT. Data visualisation method aims at creating visual representation of given high-dimensional datasets, which can facilitate the viewer's understanding of the intrinsic structure of the dataset. One of the most popular schemes is to project the high-dimensional dataset to some low-dimensional spaces, maintaining the link information between the objects in the dataset. Classical methods like Laplacian Eigenmap, Isomap, Local Linear Embedding, t-Stochastic Neighbour Embedding, etc. try to keep the local structure of the dataset, or to preserve the probability distribution of the object pairs based on their local structure. These mechanisms lack considering the relative relationship between data cohorts. COVA utilizes the idea of prototype to control the cohort position and relationship. Those prototypes can be seen as the representative of the data cohorts. COVA keeps the relationship between objects and prototypes as well as the local structure of the objects during the data embedding process. However, the local neighbour preservation rate is sacrificed. Through the experiments, we’ve noticed that the Local Ordinal Embedding method has the outstanding performance on local neighbour preservation. Based on the idea of COVA, we introduce the anchor points to the Local Ordinal Embedding method, try to get a high neighbourhood preservation score as well as preserve the relative location of cohorts.

ABSTRACT. WebAssembly (WASM) is a new Intermediate Representation (IR) designed for web execution. Its novel design as a compilation target for the web provides an opportunity for investigating features that wins its performance, and novel runtime optimisation techniques.

Currently all four major browsers (https://webassembly.org/) integrate WebAssembly within their existing JavaScript engines. This research provides a standalone WASM implementation which facilitates investigating WASM IR runtime performance as a standalone language, its interoperability with JavaScript and other JVM languages, and novel instrumentation techniques while hosted on the Java Virtual Machine (JVM).

ABSTRACT. Bayesian statistics enables us to use a prior information to control the posterior of a parameter. Such a prior can be learned via empirical Bayes, i.e., learned from data itself. Meta learning is a method allows an agent to learn some experience from task level information, such learned experience can be utilized for fast adaption to a new task and few shots learning. The research is about encoding task level information into a learned prior such that an agent can perform a better exploration strategy for structured tasks in a Bayesian fashion.

ABSTRACT. In the open science movement, there is an increasing demand to manage scientific data. This indicates an urgent need to join and share research data in order to reuse it effectively. However, the heterogenous and complexity of scientific data has led to increased requirements to reuse this data. Therefore, FAIR (Findable, Accessible, Interoperable, and Reusable) principles have been emerged to provide a collaboration environment for sharing research data. These independent principles for data management has taken hold of policy makers, funding councils, communities and publishers. However, these are principles that are a journey or set of aspirations that are now elevated to rules and metrics. Moreover, the FAIRification of processes, datasets and practices is not yet understood, and the full implications for datasets, researchers and agencies are not fully developed. In particular FAIR is not FREE. This research will examine the implications of adopting FAIR using access to the European Research Infrastructures ELIXIR and EOSC as a vehicle.

ABSTRACT. Cosmic magnetism is the study of the largest physical scale magnetic fields in the universe, and is a field set to benefit immensely from new polarisation sky surveys with telescopes such as the Square Kilometer Array. The quantity of data produced by such surveys motivates a machine learning approach. My research aims at introducing machine learning techniques to this field, particularly through using convolutional neural networks to detect 'complexity' in the structure of magnetic fields. This necessitates some focus on the issue of class imbalance in machine learning, and of dealing with the unlabelled data provided by real telescope observations.

ABSTRACT. This paper uses network motifs to investigate the behaviour of people discussing vaccinations on five bulletin board sites. We find a significantly higher degree of transitive and other structured interaction between users than occurs in networks generated at random, indicating that users tend to agglomerate in the same threads rather than spread comments over multiple threads. We also detect potential differences in the prevalence of particular network motifs according to demographic characteristics, with indications that users of a male-oriented bulletin board tend to post in fewer vaccinations threads than users of other social networks. The results demonstrate the utility of using network motifs for characterising communication patterns within social media, since the algorithms scale to large networks, and the motifs can be interpreted as units of functional behaviour encoded in the network.

ABSTRACT. The introduction of consumer-grade RGB-D (colour and depth) cameras such as Microsoft Kinect, Occipital Structure or Google Project Tango into the computer vision research community has sparked interest in developing efficient algorithms for dense 3D reconstruction. While most work has focused on modelling static environments, in recent years there have been significant advances towards real-time non-rigid reconstruction.

Current state of the art algorithms allow users to either reconstruct large, static scenes, or small dynamic scenes, such as a single person, but not both at the same time. This project addresses the issue and presents a system capable of performing live reconstruction of a non-rigidly moving subject and the surrounding static environment by semantically labelling the scene.

ABSTRACT. This poster presents an overview of my PhD research. The research involves classifying fundamental aspects (primitives) of Artificial Intelligence systems, their properties, metrics, requirements, and limitations, and creating and applying new representations (e.g. diagrammatic) in order to describe and reason about AI. It is interdisciplinary, with grounding in Artificial Intelligence, Human-Computer Interaction, Cognitive Psychology, and Semiotics. The impact is improved representation, communication and reasoning about AI systems, as well as an improved understanding about impact of AI.

ABSTRACT. Synthetic antiferromagnetic spintronics has generated much interest for device applications due to wide ranging tunability of their magnetic properties [1]. The structure of a synthetic antiferromagnet (SAF), two ferromagnetic layers separated by an ultrathin (~0.8nm) non-magnetic, here ruthenium, spacer layer. In such a stack, the magnetic properties can be tailored through the RKKY interaction with the presence of ferromagnetic or antiferromagnetic interlayer coupling dependent on the thickness of the spacer layer. In the antiferromagnetic case, similarly to natural crystalline antiferromagnets [2], much faster spin dynamics compared to the single film or ferromagnetically coupled state are displayed due to the interlayer coupling introducing additional resonant modes [3,4]. In this work, fabrication of SAFs incorporating Co0.2Fe0.6B0.2 as the ferromagnetic components has been carried out due to its proven ability to form a SAF [4]. The SAFs have been fabricated using magnetron sputtering and characterized by X Ray Reflectivity (XRR) to measure layer thickness, Vector Vibrating Sample Magnetometry (VSM) to verify basic antiferromagnetic coupling and a Vector Network Analyzer – Ferromagnetic Resonance (VNA-FMR) setup which allows key dynamic properties to be explored.

ABSTRACT. The energy for data transfer has an increasing effect on the total system energy as technology scales, often overtaking computation energy. To reduce the power of inter-chip interconnects, an adaptive encoding scheme called Adaptive Word Reordering (AWR) is proposed that effectively decreases the number of signal transitions, leading to a significant power reduction. AWR outperforms other adaptive encoding schemes in terms of decrease in transitions, yielding up to 73% reduction in switching activity. Furthermore, complex bit transition computations are represented as delays in the time domain to limit the power overhead due to encoding. The saved power outweighs the overhead beyond a moderate wire length where the I/O voltage is assumed to be equal to the core voltage. For a typical I/O voltage, the decrease in power is significant reaching 23% at just 1 mm long interconnect.

ABSTRACT. We study the problem of for-getting, also known as Uniform Interpolation, under background theoryand propose a sound and complete method in the description logicALC.To our knowledge this problem was not studied before, instead, forget-ting on a centralised ontology was the main focus of research. The aimof forgetting is to remove a set of symbols from an ontology, while pre-serving the consequences over the remaining symbols. With backgroundontology involved, we may also desire to preserve the consequences overthe reduced set of symbols together with the symbols in the backgroundontology.

ABSTRACT. Catering for ontology summary and reuse, several approaches such as modularisation and forgetting of symbols have been developed in order to provide users smaller sets of relevant axioms. We consider different module extraction techniques and show how they relate to each other. We also consider the notion of uniform interpolation that is underlying forgetting. We show that significant improvements in the performance of forgetting can be obtained by applying the forgetting tool to ontology modules instead of the entire ontology. We investigate combining several module notions with uniform interpolation and we provide a preliminary evaluation forgetting signatures based on the European Renal Association subset from SNOMED CT. We provide possible explanations for why modularity helps forgetting symbols from large-scale ontologies in practise. To facilitate the experiments, we develop a signature extension algorithm for the SNOMED CT ontology to additionally include related symbols.

ABSTRACT. Multi-cloud based Collaborative Big Data Computation using a distributed computing service are susceptible to threats and attacks from unauthorised outsiders as well as authorised insiders. As the existing security solutions are not designed for this complex setting involving entities from different domains at multiple layers, they might not be applicable to applications in this context. Designing a security solution that is effective, efficient, and scalable is still an open and very challenging research issue. This research focuses on the security solution of authentication which is a preventive security measure commonly used as the first line of defence. In this presentation, we present a novel authentication framework consisting of an entity authentication scheme and a data authentication scheme.

ABSTRACT. Third-party based mobile health monitoring systems are vulnerable to threats not only imposed by outsiders but also authorized insiders, e.g. employees of the third-party service provider. This paper examines issues in this context and proposes a novel framework, called a secure and ID privacy preserving framework (SPDC), to facilitate secure and ID privacy-preserving data collections from remote patients. The framework has a number of properties: (i) it supports three different modes of data collections and different treatment of data with two levels of urgency, (ii) it protects the confidentiality and authenticity of data being collected, (iii) it preserves patients' ID privacy against external entities as well as entities that are involved in facilitating the data collections, (iv) it uses distributed load-sharing so that no single entity is wholly responsible for the task of data collections. The ideas used in achieving these properties include the use of structured use of multiple data collection servers and multi-level hierarchical pseudonyms to hide patients' IDs as well as their communication patterns, and the separation of duties and pseudonym and blind-token based authentication methods so that patients' can be identified and authenticated without revealing their real IDs. Security analysis against design requirements and anonymity evaluation using entropy method are carried out to demonstrate that the framework can resists attacks on data security and protects the patients' IDs from being revealed to unauthorized entities.

ABSTRACT. This paper presents an idea of applying Deep Learning Algorithms on understanding Traditional Chinese Painting and trying to investigate why their outcomes have huge deviations from what we want to attain. The project focuses on both classification and style transfer areas. The final purpose is to improve Style Transfer performance on Generating Traditional Chinese Painting basing on the knowledge returned and refined during classification learning. We have introduced an Assembled Region Proposal Network which concatenates low-level visual information and high-level semantic knowledge to reduces coarseness in region-based object detection. By comparing the image-level classification performance of various detectors, we have shown that our method increases the sensitivity of small object detection. Also, it achieved at least 3% accuracy improvement comparing to the other detectors on Natural Images data. However, after running experiments on Traditional Chinese Paintings data, we have found our method's outcomes came in significantly below expectation. Simultaneously, the performance of all tested detectors in this paper has sharply declined with an accuracy drop from 10% to 17%.

ABSTRACT. The Arabic Language originally had an old fashioned form of writing which is found in historical data and the Holy Quran. These types of writing styles in Arabic texts give a degree of difficulty in segmenting the letters and reading the text. This art form is still used today for various purposes in drawing and calligraphy. Since it originated a long time ago, it is mostly reflected in Islamic culture and the use of quotes from the Holy Quran. The calligraphy shows a different representation of Arabic texts in a more cursive style and a mixture of complex constructed word forms. The challenges of this type of text motivates the discovery of a way to simplify reading and to digitalise it more easily. The study is constructed to achieve a solution for reading such calligraphies computationally and for determining the text that they represent. This research combines powerful methods in artificial intelligence. The development system is a mixture of image processing, text mining and machine learning methods.