Sydney eScholarship
Not a member yet
31878 research outputs found
Sort by
Neural Architecture Design and Compression for Efficient Vision Perception
No full textIn recent years, intelligent systems have evolved significantly, transforming daily life. To maximize their impact, efficient deployment on edge devices—smartphones, smartwatches, robots, and autonomous vehicles—is essential. Deep neural networks, foundational in computer vision, offer powerful feature encoding but demand substantial computational resources, leading to high energy consumption and carbon footprint.
This thesis focuses on developing compact yet high-precision deep learning models that balance performance and efficiency. It explores efficient vision backbones and compression techniques to support various tasks while ensuring deployability. We propose a hybrid architecture integrating transformers for global dependencies and CNNs for local feature extraction, replacing traditional components with fully connected layers to enhance efficiency. This design reduces complexity while maintaining accuracy.
We further investigate training a unified model for multiple vision tasks through a data-efficient strategy, enabling the model to handle both high- and low-level tasks. Extending to multi-modal learning, we introduce an efficient fusion framework to enhance AI perception in real-world applications. Additionally, we refine knowledge distillation for compact models, reassessing existing methods to improve real-world applicability. Specifically, for object detection, we highlight the overlooked role of background information and propose a decoupled distillation method that enhances performance.
This thesis presents practical solutions for lightweight neural networks, enabling AI deployment in resource-constrained environments. By optimizing deep learning models for efficiency, it contributes to the accessibility and sustainability of AI across various domains
Search for the lepton flavour violating decay B0 -> ell tau at Belle II
No full textToday, numerous avenues are explored to search for any evidence of new physics beyond the
Standard Model, one of which is rare particle decays. Such decays are promising as their
suppressed decay rates allow for any enhancements from new physics contributions to be more
readily seen. One such rare decay of interest is B0 → ell tau (ell = e, μ) which is a lepton flavour
violating decay making it forbidden within the conventional Standard Model framework. However,
within an extended Standard Model that includes neutrino oscillations, which there is strong evidence
for, the decay can occur but at a rate far beyond any current or future experimental sensitivity.
Therefore, observation of B0 → ell tau would be strong and clear evidence of new physics beyond
the Standard Model. This study presents the first search for B0 → ell tau using 365/fb of data
collected by the Belle II experiment from 2019-2022. The Belle II experiment is situated around the
interaction point of the SuperKEKB collider which is an asymmetric e+e− collider designed to mass
produce B mesons. The primary aim of the study presented in this thesis is to develop an analysis
technique for future hadronic tag-based B0 → ell tau searches at Belle II that yields an improved
sensitivity over previous analysis strategies employed by other experiments. This is achieved by
partitioning the tau decay into leptonic and hadronic modes, and designing separate signal extraction
methods that are optimised for each as opposed to employing a global strategy for all tau decays.
Focusing only on the leptonic tau decays for now, upper limits on the B0 → ell tau branching fraction
are derived using a hybrid cut-and-count approach based on the Rolke-Lopez interval method. The
results from just the leptonic tau decays have a sensitivity that is competitive with the current worldleading
upper limits using only a fraction of the dat
Are users ready to accept fully flexible walking in on-demand mobility?
No full textOn-demand ride pooling benefits from updating its routing decisions in real-time as new information becomes available, as well as from optimising the pickup and drop off (PUDO) points to avoid long detours. Both features indicate that it could be plausible to implement flexible walking, i.e., deciding the PUDO points dynamically, as opposed to informing users upfront. However, this could reduce the reliability of ride pooling thereby influencing the user experience. In this paper, we analyse data extracted from a labelled discrete choice experiment, consisting of three distinct ride pooling alternatives: door-to-door, fixed walking where walking time is informed upfront, and flexible walking where walking time is expressed as a time interval. Each alternative is further described by the following battery of attributes: price, in vehicle and waiting times, and emission savings compared to conventional petrol vehicles. The empirical analysis is performed via the density de-compositional version of the non parametric random effects Logit Mixed Logit model. Our main findings are as follows: i) There is a discontinuous zero-walk effect: passengers strongly prefer door-to-door services to
walking even a minimal distance; ii) Users prefer to know their walking direction, as the willingness to pay to reduce walking time is significantly higher when the PUDO point is not fixed in advance; iii) The so-called reliability ratio, which compares the value of reliability to the value of time, is approximately 0.68 for walking time — significantly greater than previous values obtained for either waiting or in-vehicle time indicating that reliability is relatively more important for walking. All of this implies that flexible walking would be desirable only if the operational benefits are very large. On the
other hand, those three findings reflect the average, but we do identify a percentage of the population willing to embrace flexible walking, suggesting that offering both fixed and flexible times to the users can be the best option
Enhancing Medical Record Comprehensibility: Using Large Language Models to Produce Simplified Narratives of Image Reports in Electronic Medical Data
No full textThis thesis investigates the use of language models in clinical applications where input documents are long and questions can be complex and require advanced reasoning. The research aims and objectives in this thesis contain two parts: one is to find and evaluate the appropriate solution that enhances the model performance under clinical settings; the other one is to find the solution to modify the models for better performance under this circumstance. Two approaches are developed to address the problem. In the first study, the RAPTOR framework extends a language model's ability to make sense of local and global information from long documents with its unique hierarchical tree structure datastore. The approach may be beneficial where cloud-based large language models (e.g. GPT-4o) cannot be used due to data privacy or reproducibility issues. Specifically, RAPTOR can be tailored to address clinical tasks including extracting critical patient information and summarizing clinical notes from long documents. In the second study, I developed and tested an optimized language model that uses a continual pre-training process to incorporate domain knowledge with a Llama-3.1-8B language model, with a novelly collected, organized and preprocessed Clinical Trial registration dataset called CiTi. The dataset contains 358870 preprocessed clinical trial registration reports and 1401401 related publication abstracts. This study aimed to develop a language model that adapts clinical trial registration data as its specialty domain and has more understanding of this domain than general large language models. The two solutions evaluated in this thesis show that with the updated configuration, it is possible to achieve state-of-the-art performance using locally implemented language models. Future research should consider how specific configurations or auto-configurations better suit simple and complex questions
Multivariate Volatility Measures and Models with Applications
No full textThis thesis explores methodologies for modelling and estimating correlation and covariance
dynamics, presenting advancements in statistical approaches and their applications across multiple
domains. We provide a comprehensive literature review of existing methodologies for modelling
covariance matrices, focusing on their advantages, limitations, and practical implications, which
highlights the need for efficient estimators and dynamic modelling techniques to address challenges
such as heteroskedasticity, non-positive definiteness, and dynamic correlation structures. With our
proposed range-based correlation matrix measures, we extend the two-stage multivariate Conditional
Autoregressive Range Model (MCARR)-return models to directly model covariance matrix series
using the Wishart distribution. Through simulation studies, we compare two approaches: modelling the covariance matrices and modelling the variances and correlation matrices. Correlation matrix
modelling demonstrates better performance, guided by specific priors
and stationary conditions
Optimising Psychotropic Medication Use in Older Adults and People with Dementia
No full textWith global population ageing, dementia has become a leading cause of disease burden in older adults. Psychotropic medications are widely used to manage mental health conditions in this group but carry significant risks, including cerebrovascular events and mortality, with antipsychotics being of highest concern. This thesis examines prescribing patterns and determinants of psychotropic use in older Australians—particularly those with dementia—and explores strategies to optimise safe, appropriate use.
Chapter 1 outlines the clinical challenges, pharmacological classes, age-related pharmacological changes, and heightened risks of drug-related problems, focusing on antipsychotic use for behavioural and psychological symptoms of dementia (BPSD). Two objectives guide the work: (1) assess psychotropic use and associated factors, and (2) investigate treatment modifiers and predictors of risperidone response in dementia.
Chapter 2 uses linked 2021 Census–PBS data for 3.85 million Australians aged ≥65, finding one-third received at least one psychotropic, with antidepressants, opioids, and benzodiazepines most common; 8.4% had psychotropic polypharmacy. Use varied significantly across dementia status, Aboriginal and Torres Strait Islander peoples, CALD groups, and non-private dwellings, highlighting inequities.
Chapter 3 analyses six clinical trials, finding risperidone effective for psychosis, aggression, and anxiety/phobias, but not for affective, activity, or sleep disturbances. Modifiers included sex, BMI, endocrine disease, and race. Early response at week 2 strongly predicted later outcomes.
Chapter 4 synthesises findings, proposing a framework for personalised antipsychotic prescribing. The thesis underscores high psychotropic use and disparities, calling for targeted, evidence-based, and culturally informed strategies to ensure safe, equitable use in aged car
The Unpopular Men of Popular Feminism: Heteropessimistic feminisms in the era of the manosphere
No full textThe Unpopular Men of Popular Feminism: Heteropessimistic feminisms in the era of the manosphere
The latest battles between popular feminism and popular misogyny in Western culture have been
characterised, since 2018, by spectacular expressions of misogyny enjoying institutional backing at
the highest levels, and the emergence of popular feminisms that incorporate increasingly right-wing
ideas and express a prevailing heteropessimism. Renewed attempts, including by feminists, to define
“woman” solely in reproductive terms aligned with opposition to what is called “gender ideology” have
gained traction across several continents (see Butler, 2024). Allied with “strong man” (and
sometimes, strong woman) politics, and part of a broad network of “reactionary digital politics,”
misogyny currently exerts considerable influence over the terrain of popular culture (see Butler, 2024;
Cappelle, 2024; Finlayson, 2021; 2022; Kay, 2024). This thesis examines how popular feminism is
contributing to this changing context and how it too has changed in recent decades.
This thesis examines the representational politics of a range of cultural phenomena, including online
masculinist communities known as the “manosphere,” transphobic feminists, and histories of Western
radical feminisms. It evaluates how the rightward cultural shift I sketched above is imbricated with the
emergence of two distinct modes of popular feminism, which I call “orthodox” and “dissident.” While
orthodox feminism represents men as patriarchal threats to the realisation of a feminist social order
and dissident feminism represents men as natural threats to women, both feminisms draw
conclusions about “man” as a social category by citing examples from the manosphere. This shared
field of reference, I argue, is both symptomatic of, and has a limiting impact on, the affects towards
men—and therefore, the politics— that are readily intelligible in and as “feminism.
Patient-Reported Outcome Measures in Cataract Surgery
No full textThis thesis investigates the implementation and applicability of Patient-Reported Outcome Measures (PROMs) in routine cataract surgery care in Australia, with a focus on improving patient-centred outcomes and aligning local practice with international standards. While PROMs are increasingly integrated into clinical workflows worldwide to guide decision-making and evaluate healthcare quality, their use in Australia remains largely confined to research and benchmarking. Through a systematic review and Rasch analysis, this work validates the psychometric robustness of the Catquest-9SF questionnaire and demonstrates its superiority over the Priquest in measuring vision-related quality of life following cataract surgery. The thesis highlights the limitations of existing PROMs in reflecting the outcomes of modern cataract procedures, such as those involving multifocal intraocular lenses and immediate sequential surgeries. It also presents a practical implementation protocol using a web-based platform to address barriers to routine PROMs uptake. The findings underscore the importance of tailoring PROMs to evolving surgical techniques and patient preferences, and advocate for their wider adoption in Australia to support better clinical practice, enhance shared decision-making, and improve surgical outcomes
Kinematics-Driven Motion Analysis for 3D Hand Trajectory Prediction in Virtual Reality
No full textHand motion plays a fundamental role in how people interact with their environment, making hand trajectory prediction and analysis vital across numerous domains, including Human-Computer Interaction (HCI) and Human-Robot Interaction (HRI). Accurate hand motion modeling and prediction are particularly beneficial in Virtual Reality (VR) applications, where they can reduce system latency and enable the development of novel and immersive experiences.
However, despite prior work on various statistical and deep learning approaches, challenges persist in developing accurate, efficient, and generalizable models for deployment in real-time applications on devices with limited resources.
This thesis addresses these challenges by introducing novel techniques for analyzing and predicting hand motion, integrating empirical data and kinematics-based approaches to enhance existing mathematical models and statistical methods.First, we conducted a user study with 20 participants performing hand movements in VR to address the limitations of existing datasets. Using the collected data, we developed user-specific predictive models tailored to individual motion patterns, achieving high accuracy in hand motion prediction. Building on these personalized models, we developed generalized models that maintain comparable performance levels. These generalized models are designed to adapt across a broader user base without requiring individual customization, balancing accuracy and generalizability.
This thesis advances the fields of HCI and HRI by deepening the understanding of human hand motion and presenting techniques for accurate and efficient motion prediction. It offers a scalable approach that balances customization with broader applicability by introducing user-specific and generalized models. The future of this research lies in expanding it to encompass the entire human body, with the potential to significantly impact areas such as VR and collaborative human-robot environments
Teaching Sexual and Reproductive Health and Rights at the University: A Toolkit for Nursing and Social Work Educators (Australian Edition)
No full textThe SWAN-SR Teaching Toolkit is a co-designed educational resource developed to support university educators in teaching sexual and reproductive health and rights (SRHR) to Nursing and Social Work students. Created through a cross-national collaboration between educators, students, and frontline practitioners in Australia and Hong Kong, the toolkit responds to identified gaps in SRHR education by offering inclusive, trauma-informed, and practice-oriented content.
It includes adaptable lesson plans, case studies, reflective exercises, and guidance to help educators facilitate safe, supportive, and person-centred learning environments. The resource promotes critical thinking, professional readiness, and inclusive care, and is intended for integration into both standalone sessions and broader curricula