DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
Chain-of-thought tuning for reducing reasoning hallucinations in MLLMs
No full textMultimodal large language models (MLLMs) have advanced vision–language understanding
but often hallucinate on questions involving multiple visual conditions.
This dissertation mitigates complex hallucinations by creating a largescale
simple–complex VQA dataset, pairing single-attribute questions with composite
queries. Grounded in MSCOCO images, the dataset undergoes rigorous
verification to ensure question-image alignment. A benchmark is proposed, with
the evaluation metrics including accuracy, precision, recall, and F1 metrics.This
work quantifies hallucination rates in three open-source models: LLaVA-1.5, InstructBLIP,
and mPLUG-Owl2 on our proposed benchmark. To reduce hallucinations,
This work propose two methods: chain-of-thought prompting, eliciting
explicit reasoning during inference; and parameter-efficient LoRA instruction
fine-tuning. Both methods substantially decrease hallucinations on composite
questions, achieving accuracy improvements of 44%–56% for LLaVA-1.5 without
afftecting the performance on simple tasks. Qualitative analysis demonstrates
improved interpretability, providing a reliable framework for evaluating and enhancing
multimodal reasoning.Master's degre
Understanding mental health help-seeking behaviours among university students in Singapore
No full textUniversity students in Singapore face a high risk for mental health distress, yet their help- seeking rates are low, with a preference for informal support. This study investigates the factors influencing professional mental health help-seeking within this population, proposing an extended model of the Theory of Planned Behaviour (TPB) that integrates cultural determinants unique to Singapore. A total of 167 respondents were recruited through convenience sampling. The findings revealed that attitudes and subjective norms significantly predicted help-seeking intentions, with subjective norms as the strongest predictor. Self-stigma influenced help-seeking attitudes, but self-reliance did not. Mental health literacy did not moderate these relationships. Implications for national mental health interventions include leveraging social norms and reframing help-seeking through mental health literacy and stigma reduction.Bachelor's degre
Nature empowered single-session therapy (NEST): a pilot study on nature-based single-session interventions for mental health
No full textMental health challenges are increasingly affecting individual and societal functioning, highlighting the urgent need for accessible and preventive interventions (Wu et al., 2023). Nature-based interventions (NBIs) have the potential to promote mental well-being but are often limited by the multi-session format (Schleider et al., 2020; Song et al., 2019). On the other hand, single-session interventions (SSIs) benefits from the brief and scalable format but with uncertain long-term benefits (Hecht et al., 2023). To address these limitations, the present study introduces Nature Empowered Single-session Therapy (NEST), a novel integration of NBIs and SSIs. This study aimed to examine both the receptiveness and psychological effects of NEST through a mixed-methods approach, combining quantitative and qualitative data. 52 undergraduate students participated in a one-hour NEST session, which included nature immersion and mindfulness elements. Quantitative findings indicated high receptiveness to NEST across acceptability, appropriateness, and feasibility dimensions. Participants also showed significant immediate reductions in perceived stress and helplessness following the session. Although follow-up analyses suggested some sustained effects on anxiety, these findings may be influenced and explained by confounding factors. Qualitative responses supported NEST’s benefits, revealing four main themes: Mindfulness and Relaxation, Connection with Nature, Awareness of Surroundings, and Psychological Benefits. These themes offer insight into potential mechanisms of change. While limited by a small sample, this study offers promising preliminary evidence for NEST as a feasible and accessible single-session intervention. Future research should aim to replicate and extend these findings in larger, more diverse populations to enhance the intervention’s long-term efficacy and reach.Bachelor's degre
Validating and categorising AI-designed anti-coronavirus peptides
No full textThe Coronavirus disease 2019 (COVID-19) pandemic, caused by Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) virus, has led to millions of deaths worldwide, straining global healthcare systems and economies. While most current therapies primarily target the receptor binding domain (RBD) of Spike 1 (S1) protein, high mutation rate in RBD leads to rapid immune escape and limits their long-term effectiveness against emerging variants. In contrast, the S2 subunit, containing fusion peptide (FP) and heptad repeat 2 (HR2), is highly conserved. This plays a critical role in mediating viral-host membrane fusion and presents promising targets for broad-spectrum antiviral therapies. This project explores the inhibitory potential of AI-designed anti-coronavirus peptides targeting the conserved regions in S2. Using biochemical and biophysical assays, fifteen FPs and four optimised HR2s peptide inhibitors were screened to assess their binding affinity and structural stability. One FP and one HR2 peptide demonstrated strong binding, with favourable dissociation constants (KD), suggesting their potential as fusion inhibitors. While AI accelerates the early stages of drug discovery, experimental validation remains essential. This project highlights a pipeline for translating AI-designed peptides into viable therapeutic candidates, offering an alternative strategy against SARS-CoV-2 and its variants by focusing on structurally conserved regions like S2.Bachelor's degre
Mechanobiological impacts of engineered nanoparticles on mammalian cell dynamics
No full textThe extensive use of nanoparticles (NPs) across various industries has sparked notable concerns regarding their environmental and biological impact. Although engineered NPs offer substantial benefits, their interactions with biological systems are complex due to their small size and high surface area-to-volume ratio. Their potential to accumulate in the environment and unintentionally enter ecosystems poses risks to both human and environmental health. There is a significant gap in understanding how low, non-toxic concentrations of NPs affect essential processes like cell migration and division, which are vital for development, wound healing, and disease progression. These processes are tightly connected through cytoskeletal structures and signalling pathways. The study thus uses biological assays and microscopy techniques to assess NP-induced changed in cellular behaviour, with a focus on three key areas: (1) materials characterization, (2) the impact of metal oxide nanoparticles on coordinated epithelial cell rotation (CECR), and (3) the effects of biopersistent nanoparticles on cytoskeletal development in dividing cells. Our findings reveal that even at low, non-toxic concentrations (1 and 10 μg mL-1), zinc oxide (ZnO) and titanium dioxide (TiO2) nanoparticles significantly disrupt CECR in micropatterned epithelial cells via distinct, nanoparticle-specific mechanisms. In dividing cells, TiO2 nanoparticles impair filamentous actin assembly through direct binding and reactive oxygen species (ROS) generation, while silica (SiO2) nanoparticles show minimal effects. These results highlight the profound influence that nanoparticles, even at non-cytotoxic concentrations, can have on cellular behaviour, particularly through ROS signalling modulation. This study advances the understanding of how NPs alter cell dynamics, providing insights that can inform the safer design and application of nanoparticles in biomedical and environmental fields.Doctor of Philosoph
Visual quality evaluation with foundation models
No full textVisual quality evaluation, aiming at examining, diagnosing, and potentially improving
the quality of visual contents, plays a significant role in the era with enormous
visual content creation, e.g. user-generated contents (UGC), or most-recently predominant
AI-generated contents (AIGC). To sufficiently serve as catalyst for content
creation, the future quality assessment systems should not only provide precise
quantitative evaluations (i.e. quality scores) for different types of visual contents,
but also possess the capability to provide finer-grained diagnosis or reasoning on its
predicted quality scores, and ultimately provide effective guidance or optimization
direction to improve visual content quality.
In this thesis, we discuss our exploration on visual quality evaluation in two parts.
For part I (Chapters 2 to 5 ), we explore the contributions that specifically target
on quality assessment for videos, before the advent of large multi-modal models
(LMMs). Specifically, we focus on three specific challenges for video quality
evaluation before the emergence of large multi-modal models (LMMs). The first
significant challenge comes from the temporal relations in videos. To solve this
challenge, we propose the DisCoVQA to model the spatio-temporal information to
accurately assess video quality. The second challenge arises from the complexity
of videos. With the extra temporal dimension, videos are with significantly larger
data compared to images, comprising a series of frames. Therefore, we design the
well-recognized FAST-VQA and DOVER series, which propose end-to-end efficient
video quality assessment methods based on sampling techniques. The last challenge
is the limited scale of datasets, which typically requires subjective annotations
from human observers (being time-consuming and expensive to acquire), and consequently
existing video quality assessment datasets are relatively small compared
to those for other video tasks. This challenge has pushed our exploration on dataefficient
video quality assessment methods. As outcomes, these explorations have
greatly advanced video quality assessment and have been widely recognized in this
field.
In part II of the thesis (Chapters 6 to 9 ), we elaborate our contributions on QFuture
series, the pioneer and the most up-to-date studies in the visual quality
evaluation area (for both images and videos) that explore emerging LMMs on
(1) zero-shot visual quality evaluation with general-purpose foundation models
(Q-Bench), (2) explainable and open-ended visual quality evaluation (Q-Instruct
& Co-Instruct), (3) more precise and robust quantitative evaluations (Q-Align).
These studies have opened a new direction for visual quality evaluation and have
inspired several other works to explore this directionDoctor of Philosoph
One by one or all together: an experience sampling study on the effectiveness of concurrent and sequential emotion polyregulation
No full textThe field of Emotion Polyregulation (Ford et al., 2019) has substantially evolved in recent years, shifting from a focus on the number of strategies used to the types and combinations of strategies employed. While growing research has emphasized the benefits of strategy synergy, no studies to date have directly examined how implementation style—whether strategies are used concurrently or sequentially—influences polyregulation effectiveness and success. To address this gap, an Experience Sampling Methodology study involving 233 participants over a seven-day period was conducted, contributing a total of 5538 assessment entries reporting their emotion regulation experiences. The findings, which aligned with prior research, showed that majority of participants predominantly responded to an emotional episode with polyregulation. Findings also revealed that individuals who predominantly engaged in polyregulation were significantly more effective at downregulating negative affect than those who predominantly used monoregulation. Furthermore, a higher frequency of sequential polyregulation use was associated with more effective downregulation of negative affect, while a higher frequency of concurrent polyregulation use was linked to greater upregulation of positive affect. Together, these results offer the first direct evidence that implementation style plays a meaningful role in shaping polyregulation effectiveness and outcomes. This study contributes to a more nuanced and dynamic understanding of emotion regulation in daily life, extending the framework proposed by Ford et al. (2019) by underscoring the importance of not only the type of strategies used, but how they are implemented as well.Bachelor's degre
Performance study of DVB-T2 digital broadcasting system over fading channels
No full textFading is a phenomenon that affects broadcasting systems in the real world and various
technologies in the DVB-T2 system such as the use of rotated constellations and LowDensity Parity-Check help to reduce the effects of the phenomenon. In this report, we
study the effects of fading on the DVB-T2 system over a range of Signal to Noise
Ratios with various orders of modulation, rotation angles and code rates. Through
MATLAB simulation of the Common Simulation Platform, we find out which settings
produce the best performance in terms of Bit Error Rate over different fading
scenarios.Bachelor's degre
Unlocking the power of SAM 2 for few-shot segmentation
No full textFew-Shot Segmentation (FSS) aims to learn classagnostic segmentation on few classes to segment arbitrary classes, but at the risk of overfitting. To address this, some methods use the well-learned knowledge of foundation models (e.g., SAM) to simplify the learning process. Recently, SAM 2
has extended SAM by supporting video segmentation, whose class-agnostic matching ability is useful to FSS. A simple idea is to encode support foreground (FG) features as memory, with which query FG features are matched and fused. Unfortunately, the FG objects in different frames of SAM 2’s video data are always the same identity, while those in FSS are different identities, i.e., the matching step is incompatible. Therefore, we design Pseudo Prompt Generator to encode pseudo query memory, matching with query features in a compatible way. However, the memories can never be as accurate as the real ones, i.e., they are likely to contain incomplete query FG, and some unexpected query background (BG) features, leading to wrong segmentation. Hence, we further design Iterative Memory Refinement to fuse more query FG features into the memory, and devise a Support-Calibrated Memory Attention to suppress the unexpected query BG features in memory. Extensive experiments have been conducted on PASCAL-5i and COCO-20i to validate the effectiveness of our design, e.g., the 1-shot mIoU can be 4.2% better than the best baseline.Agency for Science, Technology and Research (A*STAR)Published versionThis study is supported under the RIE2020 Industry Alignment Fund – Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)
Enhancing energy efficiency and PMV in multi-storey HVAC systems using PBT with multi-agent PPO
No full textHeating, ventilation, and air conditioning (HVAC) systems in buildings, due to
their high energy demand, have become a major source of global carbon emissions;
however, they are also crucial for enhancing indoor thermal comfort and
occupant satisfaction. To address this issue, this study proposes an algorithm
based on Population Based Training (PBT) and Multi-Agent Proximal Policy
Optimization (MAPPO). This method intelligently optimizes the hyperparameters
of each model and trains the MAPPO model to achieve adaptive adjustments
of the cooling setpoints, thereby optimizing both energy management and thermal
comfort. The effectiveness of the proposed method is validated through
simulations using EnergyPlus and BuildingGym. Comparative case studies indicate
that, compared with traditional rule-based control strategies, MAPPO can
achieve a 29.4% reduction in PMV or a 3.2% reduction in energy consumption,
whereas the single-agent PPO only achieves reductions of 11.4% or 1.3%,
respectively. Moreover, MAPPO is able to maintain the average room PMV
within 0.2, with setpoint stability that is comparable to that of rule-based methods.Master's degre