Engineering

Publication Search Results

Now showing 1 - 10 of 139
  • (2020) Moradi, Marjan
    Thesis
    Millimeter wave networks promise to offer ultra-fast internet download speed, but the access points or base stations must always align the beams precisely to client devices. Efficient beam alignment for mobile users therefore is considered one of the most challenging problems facing millimeter wave networks. Existing approaches that use in-band beam alignment suffers from long alignment delays and low communication performance, especially when large number of mobile clients are connect to the access point. In this research, we explore the benefits of out-of-band inaudible sound assisted beam alignment to reduce the outage probability, thereby improving the performance gain of antenna in millimeter wave beamforming. In particular, this thesis makes three fundamental contributions. First, we analytically study the beam alignment performance of 802.11ad in the presence of multiple devices while rotating with an applicable angular velocity. We come up with a probabilistic model for required number of beacon intervals to complete antenna training in multi-users scenario for 802.11ad. Second, we propose to take advantage of inaudible sound as a side channel to detect the direction of client and assist beam alignment in millimeter wave access points. Using a combination of experimental and simulation analysis of the inaudible sound spectrum available in typical mobile phones, we demonstrate that the use of 50 Hz and 50 ms sound chirps with an array of two microphones provide efficient and reliable detection of direction. Moreover, we design a filtering approach using FDM channel access to correctly assign the sound source corresponding to the estimated angle on the receiver side. Third, we conduct a comprehensive simulation in order to evaluate the performance of the proposed sound assisted beamforming on the gain of antenna. \deleted{Initially, the proposed analytical model is validated by the developed simulation platform.} We show that our proposed algorithm achieves a significant 11 dB average gain of antenna for AP with 64 antenna sectors serving 10 users moving with walking speed of two different mobility model compared with IEEE 802.11 ad. This improvement is the result of using the proposed contention-free out-of-band sound channel to remove the existing contention-based channel access for beam alignment. We believe that our findings in this thesis shed new light on the fundamental benefits of out-of-band beamforming in crowded millimeter wave network.

  • (2020) Liu, Shifeng
    Thesis
    In this thesis, we study the sequence labeling task. Sequence labeling task is to find the best pre-defined label assignment to each token given a token sequence. For example, in named entity recognition (NER), it is to identify entity mentions from text and classify them into pre-defined types. It is a prevalent and fundamental task for many applications such as information retrieval, knowledge base construction. Though various methods have proposed, there are still urgent challenges. As most methods apply machine learning techniques requiring high-quality annotated data for training, how to obtain sufficient annotation data becomes a crucial challenge. Besides, there are other challenges such as isolation of existing methods. We firstly solve the annotation data generation problem for NER in specific domains. Currently, most NER methods are supervised or semi-supervised methods, which require human annotated data. However, human annotated data is not sufficient due to the labor and time consuming. Tackling this challenge, we propose a dictionary extension method with headword based non-exact matching to generalize distant supervision. To reduce the impact of incorrectly annotation data, we apply a weighted function. We also propose a span-level model with the corresponding dynamic programming based inference algorithm. Experiments on all three benchmark datasets in different domains demonstrate that our method outperforms previous state-of-the-art distantly supervised methods. Observing the prediction results of existing methods, we then try to re-construct connections among existing methods. For NER, there are existing methods achieving decent results. However, these methods are constructed independently and none of them utilize the strength of existing methods. We propose a stacking model to utilize the strength and avoid the weakness of the existing methods. We design meta-features based on the prediction results of existing methods to capture their properties. We also introduce external knowledge for each token. Finally, we represent the superiority of our proposed method compared with existing methods with extensive experiments. In this thesis, we also extend our understanding in NER to another sequence labeling task, hypernym detection. Hypernym detection is a key step to construct ontology in knowledge base construction. Traditional methods detect hypernyms from the predicted definition sentences, which leads to error propagation. We handle hypernym detection and definition extraction simultaneously. We propose a two-phase method with a jointly neural network for both problem in phase I and a refinement model for hypernym extraction in phase II. We carefully design features for the model in phase II utilizing the results from phase I. We conduct experiments and show the effectiveness of our method on a well-known dataset.

  • (2021) Yuan, Feng
    Thesis
    Cognitive services provide artificial intelligence (AI) technology for application developers, who are not required to be experts on machine learning. Cognitive services are presented as an integrated service platform where end users bring abilities such as seeing, hearing, speaking, searching, user profiling, etc. to their own applications under development via simple API calls. As one of the above abilities, recommender systems serve as an indispensable building brick, especially when it comes to the information retrieval functionality in the cognitive service platform. This thesis focuses on the novel recommendation algorithms that are able to improve on recommendation quality measured by accuracy metrics, e.g., precision and recall, with advanced deep learning techniques. Recent deep learning-based recommendation models have been proved to have state-ofthe-art recommendation quality in a host of recommendation scenarios, such as rating prediction tasks, top-N ranking tasks, sequential recommendation, etc. Many of them only leverage the existing information acquired from users’ past behaviours to model them and make one or a set of predictions on the users’ next choice. Such information is normally sparse so that an accurate user behaviour model is often difficult to obtain even with deep learning. To overcome this issue, we invent various adversarial techniques and apply them to deep learning recommendation models in different scenarios. Some of these techniques involve generative models to address data sparsity and some improve user behaviour modelling by introducing an adversarial opponent in model training. We empirically show the effectiveness of our novel techniques and the enhancement achieved over existing models via thorough experiments and ablation studies on widely adopted recommendation datasets. The contributions in this thesis are as follows: 1. Propose the adversarial collaborative auto-encoder model for top-N recommendation; 2. Propose a novel deep domain adaptation cross-domain recommendation model for rating prediction tasks via transfer learning; 3. Propose a novel adversarial noise layer for convolutional neural networks and a convolutional generative adversarial model for top-N recommendation.

  • (2020) Kariri, Elham
    Thesis
    Mentoring is known to effectively improve professional development. The advancements in Information Technology area have positively impacted the process of mentoring through a more technology-mediated form of mentoring known as e-mentoring or online mentoring. Online mentoring had a particularly strong effect in improving the learning opportunities in online programming communities where mentees and mentors interact with each other from around the world in a mutually beneficial learning experience and collaboration. Yet, the lack of a coherent understanding of different characteristics (e.g., opportunities, challenges, activities, and strategies employed by mentees and mentors) of e-mentoring in online programming communities and lack of knowledge about mentoring aspects of applying e-mentoring in different types of online programming platforms inhibit us from an informed design or redesign of systems for e-mentoring in such communities. With a specific focus on those shortcomings, this research presents several empirical studies to advance the understanding of e-mentoring in online programming communities. First, we investigate the emerging opportunities and challenges faced by e-mentoring in online programming community. Next, we identify and classify e-mentoring activities carried out in this context. We investigate the strategies employed to overcome e-mentoring challenges in online programming communities. Finally, based on our findings, this dissertation proposes a conceptual framework for augmenting socio-technical systems with e-mentoring. The dissertation also provides comprehensive contributions that enhance the understanding of e-mentoring in online communities and provides improvement recommendations (e.g., encouraging academic members to help by offering their services to online communities as a part of their university work, using chatbots for automated responses to queries, and improving features to manage e-mentoring tasks and projects).

  • (2020) Wang, Kai
    Thesis
    Different types of networks widely exist in our lives such as social networks and biological networks. To analyse these networks and support network-based applications, mining cohesive structures is one of the most fundamental approaches which becomes increasingly popular recently. In this thesis, three cohesive-structure-based models and effective methods for network analysis are introduced. Firstly, at the motif level, we investigate the efficient butterfly counting problem on bipartite networks. Bipartite network arises naturally when modelling relationships between two different types of entities. Butterfly (i.e., a complete 2 x 2 biclique) is the smallest non-trivial cohesive structure in bipartite networks and plays a key role in many real-world applications. To efficiently count the number of butterflies in bipartite networks, instead of the existing layer-priority-based techniques, we propose vertex-priority-based algorithms with effective CPU cache optimizing techniques. Secondly, at the subgraph level, we study the efficient bitruss decomposition problem to discover the hierarchical relationships among cohesive subgraphs on bipartite networks. Here, we use k-bitruss as the cohesive subgraph model which is defined as the maximal cohesive subgraph where each edge is contained in at least k butterflies and we aim to find all the k-bitrusses for k >= 0. We first propose the BE-index which compresses butterflies into k-blooms (i.e., 2 x k bicliques). Based on the BE-index, the new bitruss decomposition algorithm is proposed, along with two batch-based optimizations. Furthermore, a progressive compression approach is devised which is more efficient when handling the edges with high butterfly counts. We show that our new algorithms significantly outperform the existing algorithms both theoretically and practically. Thirdly, also at the subgraph level, we propose a new cohesive subgraph model named radius-bounded k-core (RB-k- core) to capture cohesive subgraphs that satisfy both social and spatial constraints on geo-social networks. We use k- core to ensure the social cohesiveness and we use a radius-bounded circle to restrict the locations of users in an RB-k- core. We explore several algorithmic paradigms to compute RB-k-cores, including a triple-vertex-based paradigm, a binary-vertex-based paradigm, and a paradigm utilizing the concept of rotating circles.

  • (2022) Senanayake, Upul
    Thesis
    Decline in cognitive functions including memory, processing speed and executive processes, has been associated with ageing for sometime. It is understood that every human will go through this process, but some will go through it faster, and for some this process starts earlier. Differentiating between cognitive decline due to a pathological process and normal ageing is an ongoing research challenge. According to the definition of the World Health Organization (WHO), dementia is an umbrella term for a number of diseases affecting memory and other cognitive abilities and behaviour that interfere significantly with the ability to maintain daily living activities. Although a cure for dementia has not been found yet, it is often stressed that early identification of individuals at risk of dementia can be instrumental in treatment and management. Mild Cognitive Impairment (MCI) is considered to be a prodromal condition to dementia, and patients with MCI have a higher probability of progressing to certain types of dementia, the most common being Alzheimer's Disease (AD). Epidemiological studies suggest that the progression rate from MCI to dementia is around 10-12\% annually, while much lower in the general elderly population. Therefore, accurate and early diagnosis of MCI may be useful, as those patients can be closely monitored for progression to dementia. Traditionally, clinicians use a number of neuropsychological tests (also called NM features) to evaluate and diagnose cognitive decline in individuals. In contrast, computer aided diagnostic techniques often focus on medical imaging modalities such as magnetic resonance imaging (MRI) and positron emission tomography (PET). This thesis utilises machine learning and deep learning techniques to leverage both of these data modalities in a single end-to-end pipeline that is robust to missing information. A number of techniques have been designed, implemented and validated to diagnose different types of cognitive impairment including mild cognitive impairment and its subtypes as well as dementia, initially directly from NM features, and then in fusion with medical imaging features. The novel techniques proposed by this thesis build end-to-end deep learning pipelines that are capable of learning to extract features and engineering combinations of features to yield the best performance. The proposed deep fusion pipeline is capable of fusing data from multiple disparate modalities of vastly different dimensions seamlessly. Survival analysis techniques are often used to understand the progression and time till an event of interest. In this thesis, the proposed deep survival analysis techniques are used to better understand the progression to dementia. They also enable the use of imaging data seamlessly with NM features, which is the first such approach as far as is known. The techniques are designed, implemented and validated across two datasets; an in-house dataset and a publicly available dataset adding an extra layer of cross validation. The proposed techniques can be used to differentiate between cognitively impaired and cognitively normal individuals and gain better insights on their subsequent progression to dementia.

  • (2022) Flanagan, Colm
    Thesis
    Elements from cognitive psychology have been applied in a variety of ways to artificial intelligence. One of the lesser studied areas is in how episodic memory can assist learning in cognitive robots. In this dissertation, we investigate how episodic memories can assist a cognitive robot in learning which behaviours are suited to different contexts. We demonstrate the learning system in a domestic robot designed to assist human occupants of a house. People are generally good at anticipating the intentions of others. When around people that we are familiar with, we can predict what they are likely to do next, based on what we have observed them doing before. Our ability to record and recall different types of events that we know are relevant to those types of events is one reason our cognition is so powerful. For a robot to assist rather than hinder a person, artificial agents too require this functionality. This work makes three main contributions. Since episodic memory requires context, we first propose a novel approach to segmenting a metric map into a collection of rooms and corridors. Our approach is based on identifying critical points on a Generalised Voronoi Diagram and creating regions around these critical points. Our results show state of the art accuracy with 98% precision and 96% recall. Our second contribution is our approach to event recall in episodic memory. We take a novel approach in which events in memory are typed and a unique recall policy is learned for each type of event. These policies are learned incrementally, using only information presented to the agent and without any need to take that agent off line. Ripple Down Rules provide a suitable learning mechanism. Our results show that when trained appropriately we achieve a near perfect recall of episodes that match to an observation. Finally we propose a novel approach to how recall policies are trained. Commonly an RDR policy is trained using a human guide where the instructor has the option to discard information that is irrelevant to the situation. However, we show that by using Inductive Logic Programming it is possible to train a recall policy for a given type of event after only a few observations of that type of event.

  • (2022) Zhao, Runqing
    Thesis
    Emerging modes of air transport such as autonomous airport shuttle and air taxi are potentially efficient alternatives to current transport practices such as bus and train. This thesis examines bus shuttle service within an airport and air metro as two examples of network design. Within an airport, the bus shuttle serves passengers between the terminals, train stations, parking lots, hotels, and shopping areas. Air metro is a type of pre-planned service in urban air mobility that accommodates passengers for intra- or inter-city trips. The problems are to optimise the service, and the outputs including the optimal fleet size, dispatch pattern and schedule. Based on the proposed time-space networks, the service network design problems are formulated as mixed integer linear programs. The heterogeneous multi-type bus fleet case and stochastic demand case are extended for the airport shuttle case, while a rolling horizon optimisation is adopted for the air metro case. In the autonomous airport inter-terminal bus shuttle case, a Monte Carlo simulation-based approach is proposed to solve the case with demand stochasticity, which is then further embedded into an "effective" passenger demand framework. The "effective" demand is the summation of mean demand value and a safety margin. By comparing the proposed airport shuttle service to the current one, it is found that the proposed service can save approximately 27% of the total system cost. The results for stochastic problem suggest estimating the safety margin to be 0.3675 times of the standard deviation brings the best performance. For the second case, the service network design is extended with a pilot scheduling layer and simulation is undertaken to compare the autonomous (pilot-less) and piloted service design. The results suggest that an autonomous air metro service would be preferable if the price of an autonomous aircraft is less than 1.6 times the price of a human-driven one. The results for rolling horizon optimisation suggest to confirm the actual demand at least 45 minutes prior to departure. Based on data from the Sydney (Australia) region, the thesis provides information directly relevant for the service network design of emerging modes of air transport in the city.

  • (2021) Zhao, Benjamin
    Thesis
    Both researchers and industry have increased their employ of machine learning in new applications with the unfaltering march of the Digital Revolution. However, without complete consideration of these rapid changes, undiscovered attack surfaces may remain open that allow bad actors to breach the security of the system, or leak sensitive information. In this work we shall investigate attacks with and against Machine Learning, starting in the application space of authentication which has observed the adoption of ML, before generalizing to any ML model application. We shall explore a multitude of attacks from ML-assisted behavioral side-channel Attacks against novel authentication systems, Random Input Attacks against the ML models of biometrics, to Membership and Attribute inference attacks against ML models which find employ in Authentication among a host of other sensitive applications. With any proposed attack, there is an obligation to define mitigation strategies. This advancement of knowledge in both attacks and defenses will make the ever-evolving landscape that is our digital world more hardy to external threats. However, in the constant arms race of security and privacy threats, the problem is far from complete, with iterative improvements to be sought on both attacks and defenses. Having not yet attained the perfect defense, they are currently flawed, paired with a tangible cost in either the usability or utility of the application. The necessity of these defenses cannot be understated with a looming threat of an attack, we also need to better understand the trade-offs required, if they are to be implemented. Specifically, we shall describe our successful efforts to rapidly recover a user's secret from observation resilient authentication schemes (ORAS), through behavioral side-channels. Explore the surprising effectiveness of uniform random inputs in breaching the security of behavioral biometric models. Dive deep into membership and attribute inference attacks to highlight the infeasibility of attribute inference due to the inability to perform strong membership inference, paired with a realigned definition of approximate attribute inference to better reflect the privacy risks of an attribute inference attacker. Finally evaluating the privacy-utility tradeoffs offered by differential privacy as a means to mitigate the prior membership and attribute inference attacks.

  • (2021) Bai, Yu
    Thesis
    Machine learning algorithms usually have a number of hyperparameters. The choice of values for these hyperparameters may have a significant impact on the performance of an algorithm. In practice, for most learning algorithms the hyperparameter values are determined empirically, typically by search. From the research that has been done in this area, approaches for automating the search of hyperparameters mainly fall into the following categories: manual search, greedy search, random search, Bayesian model-based optimization, and evolutionary algorithm-based search. However, all these approaches have drawbacks — for example, manual and random search methods are undirected, greedy search is very inefficient, Bayesian model-based optimization is complicated and performs poorly with large numbers of hyperparameters, and classic evolutionary algorithm-based search can be very slow and risks falling into local optima. In this thesis we introduce three improved evolutionary algorithms applied to search for high-performing hyperparameter values for different learning algorithms. The first, named EWLNB, combines Naive Bayes and lazy instance-weighted learning. The second, EMLNB, extends this approach to multiple label classification. Finally, we further develop similar methods in an algorithm, named SEODP, for optimizing hyperparameters of deep networks, and report its usefulness on a real-world application of machine learning for philanthropy. EWLNB is a differential evolutionary algorithm which can automatically adapt to different datasets without human intervention by searching for the best hyperparameters for the models based on the characteristics of the datasets to which it is applied. To validate the EWLNB algorithm, we first use it to optimize two key parameters for a locally-weighted Naive Bayes model. Experimental evaluation of this approach on 56 of the benchmark UCI machine learning datasets demonstrate that EWLNB significantly outperforms Naive Bayes as well as several other improved versions of the Naive Bayes algorithms both in terms of classification accuracy and class probability estimation. We then extend the EWLNB approach in the form of the Evolutionary Multi-label Lazy Naive Bayes (EMLNB) algorithm to enable hyperparameter search for multi-label classification problems. Lastly, we revise the above algorithms to propose a method, SEODP, for optimizing deep learning (DL) architecture and hyperparameters. SEODP uses a semi-evolutionary and semi-random approach to search for hyperparameter values, which is designed to evolve a solution automatically over different datasets. SEODP is much faster than other methods, and can adaptively determine different deep network architectures automatically. Experimental results show that compared with manual search, SEODP is much more effective, and compared with grid search, SEODP can achieve optimal performance using only approximately 2% of the running time of greedy search. We also use SEODP on a real-world social-behavioral dataset from a charity organization for a philanthropy application. This dataset contains comprehensive real-time attributes on potential indicators for candidates to be donors. The results show that SEODP is a promising approach for optimizing deep network (DN) architectures over different types of datasets, including a real-world dataset. In summary, the results in this thesis indicate that our methods address the main drawback of evolutionary algorithms, which is the convergence time, and show experimentally that evolutionary-based algorithms can achieve good results in optimizing the hyperparameters for a range of different machine learning algorithms.