70040 research outputs found
Sort by
Visualising and Valuing Urban Agriculture for Land Use Planning: A Critical GIS Analysis of Sydney and Neighbouring Regions
Get PDFThe loss of a city’s agricultural lands due to land use change through urban development is a global problem, as local food production is an essential green infrastructure for intergenerational sustainability. Like many cities, much of Sydney’s rapid urban development occurs on land previously used for food production. Sydney has one of the highest rates of urban growth among Western cities and a planning strategy that marginalises its agricultural productivity. To better understand and advocate for Sydney’s capacity for food production we explore all available government datasets containing agricultural biophysical capacity using a critical GIS approach. Employing various spatial-data visualisations to contextualise agricultural production, we examine inherent biophysical agricultural capacity in Sydney and comparable regions along the eastern coast of NSW. Our approach interrogates the notion that Sydney’s metropolitan landscape is of low inherent biophysical quality for agriculture, thereby challenging current development and planning orthodoxy and policy. In ascertaining Sydney’s comparative capacity for agriculture we find that, despite current metropolitan planning policy, datasets reveal western Sydney is biophysically well suited for agriculture. Sydney overall is comparable to five of six other coastal regions of NSW and superior to at least two. While acknowledging metropolitan land use complexities that shape agricultural production in practice, we argue for improved critical application and contextual understanding of existing agricultural datasets to better inform future planning policy to advance regional food security and aid long-term sustainability
Fairness in NLP models
Get PDFThe advancement of Natural Language Processing (NLP) technologies has brought significant changes to various fields, raising concerns about fairness and ethics in NLP systems. This thesis delves into the technical dimensions of fairness in NLP systems, aiming to thoroughly examine, quantify, and develop effective methods to mitigate biases inherent in these technologies.
The research begins with a comprehensive review of the current state of fairness in NLP, conducting a rigorous examination of existing datasets and models to identify and characterize biases in widely used NLP models. This foundational analysis sheds light on the root causes and manifestations of biases, emphasizing their impact on model performance and outcomes.
The core of the study involves an in-depth analysis of fairness in NLP algorithms, exploring advanced machine learning techniques, model interpretability methods, and fairness-aware training approaches. The goal is to proactively address potential sources of bias and develop methods to rectify existing biases during the model development process.
Throughout the research, a focus on transparency and reproducibility is maintained, ensuring that bias detection and mitigation methodologies are accessible and comprehensible. This study aims to provide valuable insights to the technical community, assisting researchers and developers in creating more fair and unbiased NLP systems.
In summary, this thesis focuses on the technical aspects of fairness to enhance our understanding of various types of biases in NLP. It aims to offer practical solutions within the realm of technology, ultimately promoting the evolution of more reliable and unbiased language models and reducing the impact of biases in AI-assisted decision-making processes
Centralised Management System and Hot Transfer for ST-Elevation Myocardial Infarction in Western NSW: Closing the Gap in Current Models of Rural ST-Elevation Myocardial Infarction Care
No full textBackground: Rural vs metropolitan ST-elevation myocardial infarction (STEMI) patients experience delayed access to percutaneous coronary intervention (PCI). Existing New South Wales (NSW) Statewide Cardiac Reperfusion Strategy protocols provide thrombolysis and ambulance diversion for patients within 90 minutes of a PCI centre in regional and rural NSW. Rural patients presenting to non-PCI hospitals and those more than 90 minutes from PCI are not routinely, urgently, diverted under existing protocols. Method: Western NSW Local Health District, covering 250,000 km2 and a population of 278,759, implemented a centralised management system (CMS) in 2019, in partnership with NSW Ambulance, utilising existing STEMI thrombolysis protocols and extending “drip and ship” protocols for “hot transfer” of all patients to the 24/7 PCI centre, by direct ambulance diversion up to 120 minutes by road, or via multi-stage transfer by road or air, or via interhospital transfer. Data for 2 years post-CMS was compared to historical controls. Time from first clinical contact (FCC) to reperfusion, FCC to PCI centre, major adverse clinical events and percentage of patients undergoing angiography within 24 hours were compared in “medium” (90–120 minutes) and “long” (>120 minutes) transfer zones, not covered by existing protocols. Results: Outcomes were recorded for 274 patients before and 348 after CMS implementation (17% medium and 31% long transfer zones). Medium and long transfer zones had greater proportions of smokers and Indigenous patients than short transfer zones. There was significantly lower ambulance utilisation in the long (38%) compared with the short transfer zone (55%, p<0.001). In the long transfer zone, there were significant improvements in FCC to reperfusion (40 vs 48 minutes, p<0.05), FCC to PCI centre (296 vs 344 minutes, p<0.01), and angiography in 24 hours (77% vs 58%, p<0.01), with no significant differences in major adverse clinical events. Conclusions: A rural STEMI CMS, with “hot transfer”, can deliver patients from a vast geographical area directly to a rural PCI centre. Patients furthest away, with the greatest risk profile, benefit the most. Extension of this program and development of 24/7 PCI in NSW rural cardiac hubs stands to improve timely, definitive treatment, including access to angiography within 24 hours
Minimising the energy performance gap in Australia's commercial buildings: Energy modelling practice, process, and performance
No full textEvidence suggests that energy modelling with a ‘design for performance’ approach and Commitment Agreement has reduced the energy performance gap (EPG), primarily in Australia's prime office sector. Yet no research has documented the energy modelling practices and processes employed across the full spectrum of commercial building types in the country; what are the factors driving different types of energy modeling? And what are the challenges and opportunities stakeholders face? This study fills this research gap by surveying energy modellers and building operations teams, groups whose insights have been rarely considered up to now. The results show a marked difference in energy modeling practices between prime offices, mid-tier offices, and other commercial buildings. For instance, compliance-based energy modelling makes up 85% of modeling in non-office buildings, but only 53% in office buildings, and 49% in premium offices. The practice of engaging energy modellers throughout design and operational stages and designing to a specific operational energy performance target are also more prevalent in the office sector, than in non-offices. Energy modellers note how an inadequate scope of engagement and design uncertainty are the two greatest challenges they face in tackling the EPG, and call for energy monitoring and benchmarking, tuning and commissioning, better collaboration and education. The conclusions suggest that policymakers should explore mandating operational performance targets at the design stage, along with performance disclosure in operations, and encourage building owners to perform ongoing benchmarking, monitoring, and tracking of operational performance to reduce the EPG across the broader commercial sector
Best bang for your buck: Considerations for cost-efficiency in knowledge co-production
Get PDFKnowledge co-production is a key strategy of collaborative engagement between research and management to achieve better outcomes. Whereas the principles of knowledge co-production in support of evidence-informed policy-making are increasingly understood, our understanding of its cost-efficiency - putting benefits in relation to its cost - is in its infancy. Here, we approach this gap by exploring the key considerations for ensuring that the benefits of co-production processes outweigh the significant direct and indirect costs they can incur. We conceptualise a relationship between the costs and benefits of co-production, consider preconditions that affect those costs and benefits, and outline options for improving the cost-benefit relationship. Specifically, we explore how to maximise co-production efficiency for key principles underpinning effective knowledge co-production (context-based, pluralistic, goal-oriented, interactive) and illustrate this with a hypothetical case study of co-production for the use and management of an emerging small-scale fishery. To this end, we conclude by providing a series of guiding questions that practitioners of co-production can use to help ensure that the benefits outweigh the costs. Our results provide researchers and practitioners with improved understanding of the costs and benefits of co-production and encourage the consideration of cost-efficiency in the planning of participatory research. Further, by considering the costs and benefits of co-production processes we provide critical insights into how to ensure effective and efficient science-policy engagement where expectations might exceed limited resources. This includes enabling more transparent and accountable funding and engagement decisions while engaging multiple context-specific streams of policy-relevant knowledge for evidence-informed policy
What Are the Success Factors for a Just Transition in Critical Mineral Extraction? Analysis From the Lithium Triangle
Get PDFThe scramble to extract critical energy transition minerals creates risk of widespread negative human rights impacts. A just transition in the extraction of critical minerals must involve deep examination of the mine-community interface to gain a better understanding of the drivers of successful engagement between mining companies and communities. Drawing on fieldwork in South America's lithium triangle, this paper finds that the nature of the corporate-community relationship is increasingly key to enabling a just transition whereby communities participate in the benefits of extraction with negative impacts mitigated. It establishes that key success factors are related to empowerment of Indigenous communities and have the potential to maximise positive outcomes for communities in the context of lithium extraction. Governments and companies must embed a more bottom-up process with an end goal of communities themselves defining the parameters of what a just transition means in the critical minerals context
Exploring Language and Culture: Insights from the CogSCAN Online Survey Study
Get PDFOlder adults from culturally and linguistically diverse (CALD) backgrounds are underrepresented in, and often excluded from, dementia research. Further, language and cultural factors are known to influence performance on cognitive tests used to assess mild cognitive impairment and dementia. A new measure, the Characterising Language Experience and Acculturation Questionnaire (CLEAr-Q) was developed to measure these factors. This report is a plain language summary of the results from the CogSCAN CLEAr-Q Online Survey Study drafted with consultation from the CogSCAN Community Working Group
Multifaceted Relational Models: A Biologically Inspired Approach to Speech Recognition
Get PDFThe human brain employs a multifaceted relational thinking process for speech recognition. During this process, the human brain is likely to perceive speech in a stochastic manner. This inherent randomness enhances its ability to adapt to new inputs. Additionally, the human brain is able to identify meaningful relationships between different components of perceived speech, and further comprehend their significance for accurate recognition. For example, when distinguishing "to gather" from "together", the human brain discerns an essential relational pattern—the co-occurrence of the phone [uh] corresponding to "o" and the phone [ae] corresponding to "a". Moreover, the human brain is likely to first recognize basic phones and larger linguistic units (such as diphones and syllables) before generating the final textual transcription (i.e., a sequence of words). This process involves interactions among linguistic units at multiple scales, which collectively contribute to accurate recognition of the perceived speech.
In contrast, most prevailing artificial speech recognition systems draw limited inspiration from the human brain's functional mechanisms for speech processing. Instead, they are largely confined to architectural designs aimed at achieving a direct mapping from speech signals to textual transcriptions. Lacking stochastic modeling, these systems generally do not exhibit sufficient robustness to unseen data. Furthermore, they primarily rely on attention mechanisms to extract speech representations for recognition tasks. Using the aforementioned example, the attention mechanism only independently evaluates the occurrence of [uh] for "o" and [ae] for "a". If it detects the occurrence of either phone and assigns it a high activation value, this can result in a strong likelihood of recognizing "to gather". However, even when the input speech actually corresponds to a different phrase, e.g., "together", a highly activated [uh] for "o" alone can still cause the attention mechanism to misinterpret the input as "to gather". Failing to capture the more intrinsic relational pattern, the attention mechanism potentially leads to incorrect recognition results. Additionally, current systems do not model speech recognition as a holistic process involving interactions among linguistic unit at various scales, thereby overlooking the crucial intrinsic relationships significant for accurate recognition. Having neglected the roles that the sophisticated biological mechanisms play, artificial systems still lag behind human cognitive capabilities in speech recognition.
This thesis aims to explore the incorporation of various biological mechanisms involved in the human brain’s multifaceted relational thinking process, which are essential for human speech recognition, into artificial speech recognition systems. By doing so, it seeks to bridge the performance gap between artificial technologies and human cognitive capabilities. Specifically, a relational thinking based stochastic modeling approach is proposed to capture the inherent relational information across multiple domains of speech input, thereby obtaining improved speech representations that enhance the accuracy and robustness of downstream recognition tasks. Furthermore, a cross-layer relational thinking modeling approach is proposed, emulating the brain’s multi-scale, holistic speech processing flow by extracting relational information related to linguistic units of various scales and concurrently utilizing this information to improve the ultimate recognition task. In addition, this thesis includes experimental analyses to evaluate the impact of modeling these biological mechanisms within artificial systems. The proposed modeling techniques show effectiveness in enhancing the performance of artificial speech recognition systems, and the in-depth investigations provide insights interpreting how the modeled mechanisms aid the recognition tasks
PHOTO-CROSSLINKABLE SILK HYDROGELS FOR BIOMEDICAL APPLICATIONS
Get PDFCovalently crosslinked silk fibroin hydrogels have gained popularity over their physically crosslinked (via beta-sheet formation) counterparts due to their elastomeric nature, transparency, and ability to support cell encapsulation. Covalent crosslinking occurs through di-tyrosine bond formation between tyrosines natively found in silk. This reaction can be mediated via enzymatic-, Fenton reaction-, and photocrosslinking-based approaches. These hydrogels, however, can undergo spontaneous beta-sheet formation post-crosslinking, leading to a rapid shift from soft elastomeric to stiffer gels. This transition mechanism remains poorly understood and hampers the ability to control and use these hydrogels in biomedical applications. Silk photocrosslinking using photoinitiators ruthenium and sodium persulfate (Ru/SPS), unlike enzymatic crosslinking horseradish peroxidase and H2O2), supports rapid gel formation and is compatible with high cell-density biofabrication. Interestingly, photocrosslinking also inhibited/delayed beta-sheet formation, an effect found to be highly dependent on gelation conditions. This thesis explored the factors that influence the dynamic changes in photocrosslinked silk hydrogels and developed three diverse strategies for controlling the physical properties of silk hydrogels and spontaneous beta-sheet formation. These include: (1) control over gelation conditions, where decreasing silk concentration and molecular weight, while increasing Ru/SPS reduces the propensity for beta-sheet formation; (2) generation of silk-tropoelastin hybrids, where surface charge interactions and hydrophobic attraction between silk and tropoelastin not only limit beta-sheet formation resulting in less dynamic stiffening process, but support improved cell interactions; and (3) incorporation of calcium ions into silk hydrogels, which limits beta-sheet formation through ionic interactions and generates tough and elastic hydrogels for use in sensing and wearables applications
Presentation and Clinical Course of Leptospirosis in a Referral Hospital in Far North Queensland, Tropical Australia
Get PDFThe case-fatality rate of severe leptospirosis can exceed 50%. This retrospective cohort study examined 111 individuals with laboratory-confirmed leptospirosis admitted to Cairns Hospital, a referral hospital in tropical Australia, between January 2015 and June 2024. We examined the patients’ demographic, clinical, laboratory and imaging findings at presentation and then correlated them with the patients’ subsequent clinical course. Severe disease was defined as the presence of pulmonary haemorrhage or a requirement for intensive care unit (ICU) admission. The patients’ median (interquartile range) age was 38 (24–55) years; 85/111 (77%) were transferred from another health facility. Only 13/111 (12%) had any comorbidities. There were 63/111 (57%) with severe disease, including 56/111 (50%) requiring ICU admission. Overall, 56/111 (50%) required vasopressor support, 18/111 (16%) needed renal replacement therapy, 14/111 (13%) required mechanical ventilation and 2/111 (2%) needed extracorporeal membrane oxygenation. Older age—but not comorbidity—was associated with the presence of severe disease. Hypotension, respiratory involvement, renal involvement and myocardial injury—but not liver involvement—frequently heralded a requirement for ICU care. Every patient in the cohort survived to hospital discharge. Leptospirosis can cause multi-organ failure in otherwise well young people in tropical Australia; however, patient outcomes are usually excellent in the country’s well-resourced health system