69238 research outputs found
Sort by
The effect of short-term microgravity and hypergravity on eyelid and brow position
No full textPurposeSpaceflights induce periocular facial changes which could contribute to ocular conditions which affect astronauts. This study is to validate parabolic flight as a suitable model for studying gravitational facial changes.MethodsHealthy participants (n = 13; 37 ± 10 years) underwent short-term exposure to microgravity and hypergravity during parabolic flight. Facial images were captured and differences in measurements from pupil center to upper and lower eyelid margins (MRD1 and MRD2), and to the inferior and superior eyebrow margin (PTBi and PTBs) under normogravity, microgravity, and hypergravity were compared. A repeated measures ANOVA with Bonferroni-Holm corrected post-hoc paired t-test was used for statistical analysis.ResultsOne hundred and twenty-seven images (44 normal gravity, 43 microgravity, 40 hypergravity) revealed that short-term microgravity induced a mean PTBi increase of 2.3 mm (p s increase of 2.4 mm (p p p = 0.41). Short-term hypergravity did not change PTBs, PTBi or MRD2 but reduced mean MRD1 by 0.7 mm (p ConclusionsShort-term microgravity, similar to spaceflight, significantly elevates PTB but not MRD1. It was also shown that MRD2 was reduced. Short-term hypergravity induces significant reductions only in MRD1. Phenomena are likely explicable by co-activation of the muscles raising the eyelid and eyebrow. Comparison to space data suggests that parabolic flight offers a valid model to study periocular facial changes in microgravity
Investigating Stereotypical Bias in Large Language and Vision-Language Models
No full textFairness and bias have become critical concerns as machine learning models, including
both large language models and vision-language models, find increasing real-world applications.
While these models have enabled substantial performance gains across diverse
tasks, they typically inherit social biases from their training data, risking unfair outcomes
in areas ranging from textual analysis to image-based decision support. Existing
bias measurement techniques for large language models often rely on curated prompt
templates or global stereotype datasets, which struggle to capture regional or culturally
specific expressions and fail to address vocabulary that falls outside the model’s training
scope. Likewise, most debiasing methods for vision-language models focus on face-centric
data or assume fixed associations between text and image features, limiting their generalizability
to more diverse, real-world multimodal settings. This thesis investigates bias
from two perspectives: measuring bias in large language models and debiasing visionlanguage
models. Then, we apply both work threads to a unique multimodal dataset of
New Zealand youth well-being named Our Voices.
First, we quantify how large language models may exhibit local, region-specific biases.
We propose a LIBRA (Local Integrated Bias Recognition and Assessment) Framework
that automatically constructs culturally sensitive test sets from extensive corpora, capturing
demographic terms and regional vocabulary that a model may misinterpret or treat
stereotypically. To assess bias in the New Zealand context, we constructed a dataset of
over 360,000 test cases using New Zealand news media data. Our method then evaluates
biases by comparing the likelihood of stereotypical and anti-stereotypical outputs,
considering whether the model can accurately understand local expressions. Experiments
on New Zealand and other context corpora demonstrate how even top-performing language
models may systematically disadvantage certain social groups if the local context
is underrepresented in their training data.
Second, we shift attention to vision-language models, introducing a novel debiasing
approach grounded in causal inference called CABIN (Causal Adjustment Based INtervention).
We model the inference of the model as a causal model and identify the sensitive
attribute as a confounding factor. To estimate the confounding effect, we develop a
lightweight “mapper” network that learns transformations from text embedding to image
embedding, thus allowing us to change features about sensitive attributes in image embedding
by text. Then, we use backdoor adjustment to remove confounding influences.
Through extensive evaluations of standard benchmarks, our method demonstrates the
ability to reduce demographic skew while minimising the impact on model accuracy.
Third, we leverage the Our Voices dataset, a multimodal collection of text and images
that indicate well-being from young New Zealanders, to unify insights from both large
language and vision-language models. We use the Our Voices data as a corpus to construct
datasets and measure bias in large language models. Then, we apply our debiasing
framework to vision-language models that perform text-to-image retrieval on the data.
The results reveal how the Our Voices dataset differs from other data on the Internet,
providing a reference on fairness for analysing Our Voices data using machine learning
models
Towards a Trustworthy and Decentralised Future Internet of Things
No full textIn recent years the Internet of Things (IoT) continues to grow in both scale and importance, from approximately 16 billion devices in 2023 to a predicted 44 billion devices by
2027 [1]. This growth of the IoT o ers great opportunities to improve our understanding of and interaction with the physical world. However, in the existing IoT devices
are siloed and isolated within networks and vendor infrastructure, failing to realise the opportunities of pervasive computing and resulting in a fragmented IoT environment
where device vendors are forced to become cloud service providers and users are left with limited control over their devices and data.
In this thesis, we present an alternative future for the IoT, identifying key challenges in the creation and operation of IoT devices and systematically addressing these to create
a holistic decentralised architecture for the IoT. Beginning with a novel shared platform to support trustworthy and end-to-end secure decentralised applications (DSF), then
building on this to address the need for e cient and unambiguous speci cations for describing, (locally) discovering, and interacting with IoT services to create a decentralised
platform for IoT devices (DSF-IoT). We then extend this work to support constrained networks and devices while retaining the trust and privacy properties of the DSF-IoT,
enabling transparent interactions with and between constrained and unconstrained devices to create the rst truly heterogeneous decentralised and zero-trust IoT. Following
this we introduce decentralised registries (DSRs), expanding on prior works to enable the trustworthy global sharing and discovery of IoT services. Finally, we evaluate our
decentralised approach against existing centralised architectures, highlighting the novelty and utility of our holistic and zero-trust architecture to support future IoT devices.
Together our work provides an end-to-end alternative to the way we currently design and deploy IoT devices, decoupling devices from supporting infrastructure in order to
free device vendors from the burden of creating and maintaining infrastructure while allowing users to control the operation of their devices and the use of their data. This
work paves the way for a more trustworthy, interoperable, and human-centric IoT, aligning with the vision of pervasive computing where technology empowers people to
better achieve their goals
Early Nutrition and Resting-State Functional Connectivity in Moderate-to-Late Preterm Infants – DIAMOND Trial
No full textModerate to late preterm (MLP) infants (born at 32 < 37 weeks gestation) are understudied compared to very (28 < 32 weeks) and extremely (< 28 weeks) preterm infants but are still vulnerable to adverse neurodevelopmental outcomes. Preterm infants typically have underdeveloped brains with immature structural and functional organisation, delaying key neural processes, including the development of resting-state networks such as the default mode (DMN) and dorsal attention networks (DAN). Early nutritional strategies may improve brain maturation following preterm birth. This study explored the effects of three such interventions on within-network and between-network connectivity in the DMN and DAN in a subset of 37 MLP infants from a larger study on nutrition (the DIAMOND trial). Our subset had MRI scans at preterm and term-equivalent age (TEA). The interventions were: parenteral nutrition vs. dextrose-only; milk supplement vs. breast milk; and taste and smell of milk before gastric tube feed vs no taste and smell. We tested three primary hypotheses: 1) parenteral nutrition compared to dextrose will result in enhanced DMN connectivity, 2) receipt of formula milk supplement compared to mother's milk will alter the negative correlation between DMN-DAN connectivity, and lastly, 3) exposure to taste and smell before feeds compared with no exposure will enhance DAN connectivity. Primary analysis focused on the main effects of time, intervention group, and the main effect of their interaction. We analysed within- and between-group differences to measure three connectivity matrices: DMN within-network, DMN-DAN between-network, and DAN within-network. Our findings revealed no differences in the connectivity measures between treatment groups at TEA and no differences across time for the first two interventions. The no-exposure group demonstrated a significant difference between the treatment groups at the first scan, which did not persist at TEA. These results suggest that early nutrition does not impact the development of rs brain networks in MLP infants but requires further investigations in a larger sample
Charities no place for politics
No full textIn late December 2024 we learned that, after a four-year battle with the Charities Services, Te Whānau O Waipareira Trust looks set to be deregistered as a charity. It’s crucial that any potential decisions around deregistering the Waipareira Trust maintain the integrity of the charitable sector – otherwise trust is easily eroded
The Benefits of Being Uncertain: Perplexity as a Signal for Naturalness in Multilingual Machine Translation
No full textModel-internal uncertainty metrics like perplexity potentially offer low-cost signals for Machine Translation Quality Estimation (TQE). This paper analyses perplexity in the No Language Left Behind (NLLB) multilingual model. We quantify a significant model-human perplexity gap, where the model is consistently more confident in its own, often literal, machine-generated translation than in diverse, high-quality human versions. We then demonstrate that the utility of perplexity as a TQE signal is highly context-dependent, being strongest for low-resource pairs. Finally, we present an illustrative case study where a flawed translation is refined by providing potentially useful information in a targeted prompt, simulating a knowledge-based repair. We show that as the translation’s quality and naturalness improve (a +0.15 COMET score increase), its perplexity also increases, challenging the simple assumption that lower perplexity indicates higher quality and motivating a more nuanced view of uncertainty as signalling a text’s departure from rigid translationese
A Zoom of One’s Own: A Cyborgian Feminist Experience of Slow Academia in Academic Women’s Writing
No full textThis article considers our experiences of writing – and writing together – as academic women in post-pandemic times. We begin by acknowledging some of what we have learned from past work on slow academia and on academic women and their writing. We then introduce two theoretical lenses – Virginia Woolf’s essay A Room of One’s Own and Donna Haraway’s conceptualization of cyborg writing – to consider our shared Zoom writing, engaging with the discursive, material, and political dimensions and considering what might be meant by “slowness” in academic contexts such as ours. While Virginia Woolf provides insight into the necessary preconditions for academic women to write, our use of Donna Haraway’s cyborg writing allows a textured and spacious engagement with the complexities, multiplicities, and ambivalences of creating digital rooms of our own where communal writing and thinking can (and sometimes cannot) happen. In locating our present experience of academic work in an ongoing continuum of historical shifts and trends, we seek to articulate a feminist, empowered form of slowness that reflects the academia of our immediate present, rather than that of pre-pandemic times or of the academic generations who came before us. By interspersing the article with snippets of our mundane everyday realities as we come together to write, both online and in person, we attempt to bring into visibility the university worlds that we currently inhabit, and bring into existence the university worlds that we wish to dwell in the future
A Privacy Trojan Horse? Consumer Loyalty Programmes in the Grocery Sector and Data Privacy
No full textLoyalty schemes have become an established commercial feature in the grocery sectors of New Zealand and Australia. Ownership and operation of the schemes has recently undergone change but the essential features are likely to remain. However, there has been little scrutiny until recently of the implications for consumers of the use of personal information derived from the schemes as well as to who benefits from them. This article examines the data privacy practices of the major schemes, comparing them with similar arrangements existing in Australia, drawing on their website terms and privacy policies as well as critical scrutiny of them by the consumer and competition regulators in both countries. The article reveals an extensive data gathering and sharing ecosystem facilitated by the loyalty schemes involving many parties, including the grocery retailers but not limited to them. Significant legal and ethical issues are documented, including lack of clarity, manipulation of consumer behaviour and privacy risks from collection and profiling of individuals. Solutions through best practice are advanced through suggesting a code of practice aimed at the secto
Secluded Nature: Revitalising the Urban Fabric to Optimise Health and Well-being
No full textAs urban cities continue to grow in size and density, opportunities to
find respite from daily stress are increasingly scarce. Spending time
in nature has been proven to reduce stress and provide physiological
benefits, which prompts the question: how can we design urban spaces
to reconnect us with nature and positively impact our health and wellbeing?
This thesis integrates principles of biophilia, phenomenology, and
urban design to explore both theoretical and practical approaches to
revitalising highly urbanised environments. Central to this exploration
are the Japanese practice of Shinrin-yoku and the biophilia hypothesis,
which highlights nature’s healing power. Our innate affinity for
nature and the evident physiological benefits of immersing oneself
in experiences of nature can be harnessed and integrated into urban
settings to enhance physical and mental well-being. Architectural
apertures serve as crucial mediators between people and nature,
creating a dynamic phenomenological relationship between built
and unbuilt realms. The journey into nature is experienced through a
sequence of spatial transitions, atmospheric qualities, and moments
of pause. By amalgamating architectural apertures with nature, the
goal is to create an immersive healing journey even within dense
urban settings. To fully reconnect with the natural landscape, urban
geography should be reimagined to reflect and incorporate elements of
its original ecological character. Stream daylighting, which involves
restoring natural waterways buried beneath urban infrastructure, is
highlighted as a key method for reconnecting cities with their natural
landscapes.
The design project aims to revitalise Auckland’s City Centre by
daylighting Wai Horotiu, the buried stream running underneath the
city, to connect urban dwellers with nature’s healing power. A sequence
of interventions will transform unproductive car parks into flourishing
green spaces, offering the community a range of amenities and
atmospheric experiences. The concept of ‘secluded nature’ is central
to the project, providing an escape from daily stresses and immersing
people in intimate journeys of healing, tranquillity, and vitality
Cognitive Resources in Students’ Reasoning: Submicroscopic Representations of Redox Reactions
No full textRedox is widely recognised as one of the most challenging topics in chemistry education because it requires students to connect macroscopic, submicroscopic, and symbolic representations simultaneously. A key difficulty lies in identifying reacting species and spectator ions as well as their structure, and visualising electron transfer, an invisible and abstract process that distinguishes redox from other chemical phenomena. To address this, animations or dynamic visualisations have been frequently used to represent submicroscopic redox processes, particularly electron transfer. However, research consistently shows that even after engaging with such visual tools, students continue to struggle to understand redox. This persistent struggle suggests that the challenge may not lie solely in the quality of the visualisations, but also in how students interpret and make sense of them. Drawing on the cognitive resources and conceptual adherence notion, meaning that students could hold multiple ideas of a concept, with some ideas being held more strongly than others, this study investigates what kinds of cognitive resources students activate and how they coordinate these resources when they interpreted visual presentations of redox reactions at the submicrosopic level. Analysis of 15 first-year university chemistry students’ written responses to diagram-based questions and semi-structured interviews with six of them revealed that students not only held multiple cognitive resources, but (1) they also adhered to two cognitive resources when making sense of redox at the submicroscopic level: the ion-pair and ion-attraction ideas. The analysis also (2) identified specific contextual features that activated these ideas, particularly the type of redox reaction (metal or halogen displacement) and the nature of the prompts. Finally, (3) students were found to activate and coordinate these cognitive resources in varied ways, influencing how they interpreted electron transfer and evaluated the animations. A scientific understanding of redox involves the integration of relevant and accurate ideas, and students’ incorrect responses often reflected adherence to less relevant ideas, rather than a complete lack of understanding. The findings offer implications for animation design, teaching strategies, and future research in chemistry education