DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
Novel XAI algorithm development for analyzing machine learning models
No full textOver the years, explaining machine learning (ML) predictions has become crucial as these ML
models are increasingly being deployed in high-stakes domains such as
healthcare and autonomous driving. In Explainable Artificial Intelligence (XAI), SHapley Additive exPlanations (SHAP) is the widely used method for model
interpretability. However, despite its high usage in the realm of XAI, it
fails to differentiate between causality and correlation, often misattributing
feature importances when features are highly correlated. To address this
issue, we propose Causal SHAP, a novel framework that integrates
causal relationships into feature attribution, while preserving all key
properties of SHAP. In the framework, we combine the Peter-Clark (PC)
algorithm for causal discovery and the Intervention Calculus when the DAG is Absent (IDA) algorithm for causal strength quantification to address the weakness
of SHAP. Specifically in result, Causal SHAP reduces attribution scores
for features that are merely correlated with the target variable, and redistribute attribution scores to features that causally contribute to the target.
The effectiveness of the framework is validated through experiments on both synthetic and real-world dataset, focusing on the healthcare domain. To summarize, this study contributes
to the field of XAI by providing a practical framework for causal-aware
model explanations. Our approach is particularly valuable in domains
such as healthcare, where the data are highly correlated, and understanding
the true causal relationships is critical for informed decision-making.Bachelor's degre
"Water has memory!": Liminality and accountability in Disney's Frozen II and Moana
No full textThis thesis examines the Disney films Frozen II and Moana, interrogating how Elsa and Moana’s heightened character complexity – as a result of the postfeminist context they are produced in – manifests in their liminality. Through analysing the similarities and differences in the films’ plot structures, narrative meaning, as well as visual and auditory effects, this thesis aims to destabilise the established dichotomies present in both films, thereby creating an ambiguous middle-ground constituted by a multiplicity of narratives. It follows that Elsa and Moana’s inclination towards embracing conflicting narratives and opposing viewpoints proves their ability to effectively negotiate this ambiguity. This thesis then inspects the significance of their liminality as deeply interwoven into their identity, illuminating how both protagonists are inherently responsible for using their liminality to rectify historical injustices and restore balance. It is only in remedying injustices brought about by the wilful prioritization of subjective narratives that the protagonists ensure the continued survival of their communities.Bachelor's degre
Mixed-gradients distributed filtered reference least mean square algorithm – a robust distributed multichannel active noise control algorithm
No full textDistributed multichannel active noise control (DMCANC), which utilizes
multiple individual processors to achieve a global noise reduction performance
comparable to conventional centralized multichannel active noise control
(MCANC), has become increasingly attractive due to its high computational
efficiency. However, the majority of current DMCANC algorithms disregard the
impact of crosstalk across nodes and impose the assumption of an ideal network
devoid of communication limitations, which is an unrealistic assumption.
Therefore, this work presents a robust DMCANC algorithm that employs the
compensating filter to mitigate the impact of crosstalk. The proposed solution
enhances the DMCANC system's flexibility and security by utilizing local
gradients instead of local control filters to convey enhanced information,
resulting in a mixed-gradients distributed filtered reference least mean square
(MGDFxLMS) algorithm. The performance investigation demonstrates that the
proposed approach performs well with the centralized method. Furthermore, to
address the issue of communication delay in the distributed network, a
practical strategy that auto-shrinks the step size value in response to the
delayed samples is implemented to improve the system's resilience. The
numerical simulation results demonstrate the efficacy of the proposed
auto-shrink step size MGDFxLMS (ASSS-MGDFxLMS) algorithm across various
communication delays, highlighting its practical value.Ministry of Education (MOE)This work was supported by the Ministry of Education, Singapore, through Academic Research Fund Tier 2 under Grant MOE-T2EP20221-0014 and Grant MOET2EP50122-0018
Generalized Wigner-Smith analysis of resonance perturbations in arbitrary Q non-Hermitian systems
No full textPerturbing resonant systems causes shifts in their associated scattering poles in the complex plane. In a previous study [N. Byrnes and M. R. Foreman, arXiv:2408.11360], we demonstrated that these shifts can be calculated numerically by analyzing the residue of a generalized Wigner-Smith operator associated with the perturbation parameter. In this work, we extend this approach by connecting the Wigner-Smith formalism with results from standard electromagnetic perturbation theory applicable to open systems with resonances of arbitrary quality factors. We further demonstrate the utility of the method through several numerical examples, including the inverse design of a multilayered nanoresonator sensor and an analysis of the enhanced sensitivity of scattering zeros to perturbations.Ministry of Education (MOE)Nanyang Technological UniversityPublished versionN.B. was supported by Singapore Ministry of Education Academic Research Fund (Tier 1) Grant No. RG66/23. M.R.F. was supported by funding from the Institute for Digital Molecular Analytics and Science (IDMxS) under the Singapore Ministry of Education Research Centres of Excellence scheme and by Nanyang Technological University Grant No. SUG:022824-00001
16-bit high performance low power CMOS multiplier IC design
No full textThe growing computational demands of edge AI applications necessitate multiplier architectures that simultaneously achieve high performance and low power consumption. This paper presents a special Booth-Wallace multiplier design that successfully fulfills these dual objectives through innovative circuit-level optimizations. Employing Radix-4 Booth encoding and Wallace tree compression techniques, the proposed architecture incorporates several key low-power design strategies including module data reuse, optimized sign-bit encoding, and an improved 4:2 compressor structures. The design has a pipeline that maintains high-performance operation while effectively minimizing dynamic power dissipation.
Implemented in Verilog HDL and verified through VCS+DC synthesis, this high-performance low-power multiplier demonstrates significant improvements over traditional design. The optimized architecture achieves a 112.8% frequency increase from 235MHz to 500MHz while simultaneously reducing power consumption by 10.2% to 1.7913mW. This performance enhancement is attained with a controlled 31% area increase, representing an optimal balance for power-constrained edge computing applications. Experimental results confirm that the proposed techniques effectively address the fundamental challenges of maintaining high-speed operation while reducing power consumption in modern multiplier designs.Master's degre
Few-shot learning (FSL) for classification of tropical rainforest species
No full textForest biodiversity monitoring and conservation efforts critically depend on accurate tree species identification. Current approaches typically require large annotated datasets for effective training, creating a significant bottleneck in their deployment. This research addresses this challenge by investigating few-shot learning approaches, specifically Prototypical Networks (ProtoNets). These approaches enable models to learn from limited examples per class, dramatically reducing the annotation burden.
A critical aspect of our research is the development of effective Out-of-Distribution (OOD) detection, which enables the identification of novel species requiring expert annotation, further optimizing the use of limited annotation resources and potentially leading to the discovery of previously unidentified species. This combined approach has resulted in our proposed OOD-ProtoNets algorithm
Do not deepfake me: privacy-preserving neural 3D head reconstruction without sensitive images
No full textWhile 3D head reconstruction is widely used for modeling, existing neural reconstruction approaches rely on highresolution multi-view images, posing notable privacy issues.
Individuals are particularly sensitive to facial features, and
facial image leakage can lead to many malicious activities,
such as unauthorized tracking and deepfake. In contrast, geometric data is less susceptible to misuse due to its complex
processing requirements, and absence of facial texture features. In this paper, we propose a novel two-stage 3D facial
reconstruction method aimed at avoiding exposure to sensitive facial information while preserving detailed geometric
accuracy. Our approach first uses non-sensitive rear-head images for initial geometry and then refines this geometry using processed privacy-removed gradient images. Extensive
experiments show that the resulting geometry is comparable
to methods using full images, while the process is resistant to
DeepFake applications and facial recognition (FR) systems,
thereby proving its effectiveness in privacy protection.Ministry of Education (MOE)Agency for Science, Technology and Research (A*STAR)Submitted/Accepted versionThis work was supported in part by the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RT19/22) and 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)
Near-infrared bioorthogonally activatable fluorescence probe for in vivo imaging of immune checkpoint in cancer
No full textReal-time in vivo imaging of immune checkpoints is important for the guidance and prognosis of cancer immunotherapy. Although activatable optical probes have the advantages of high specificity and sensitivity as compared with “always-on” probes, it is nearly impossible to adopt the reactivity-based design approach for the detection of checkpoint proteins because they generally lack enzymatic activity. Herein, bioorthogonal-reaction-enabled fluorescence turn-on detection of immune checkpoint in cancer is reported. This approach involves a bioorthogonally activatable near-infrared fluorescence probe (BAP) and a transcyclooctene (TCO) tagged immune checkpoint antibody (αPDL1TCO). BAP is a hemicyanine fluorophore whose hydroxyl group is caged by tetrazine. Upon reaction with the TCOs of αPDL1TCO, the tetrazine moiety of BAP is cleaved to release the uncaged hemicyanine with a fluorescence turn-on response. BAP not only allows to specifically detect and track the fluctuation of PDL1 expression level in a murine colon cancer model during therapy but also shows a higher signal-to-background ratio than the “always-on” fluorophore conjugated-antibody, and a higher detection sensitivity than flow cytometric analysis of biopsied tumor tissues. It is expected that the design strategy of this bioorthogonally activatable probe can be applied for specific detection of other disease-related protein biomarkers without enzymatic reactivity.National Research Foundation (NRF)Ministry of Education (MOE)Submitted/Accepted versionK.P. thanked Singapore National Research Foundation (NRF) (No. NRF-NRFI07-2021-0005) and the Singapore Ministry of Education AcademicResearch Fund Tier 2 (Nos. MOE-T2EP30220-0010 and MOE-T2EP30221-0004) for financial support
Leveraging AI-generated emotional self-voice to nudge people towards their ideal selves
No full textEmotions, shaped by past experiences, significantly influence decision-making and goal pursuit. Traditional cognitive-behavioral techniques for personal development rely on mental imagery to envision ideal selves, but may be less effective for individuals who struggle with visualization. This paper introduces Emotional Self-Voice (ESV), a novel system combining emotionally expressive language models and voice cloning technologies to render customized responses in the user’s own voice. We investigate the potential of ESV to nudge individuals towards their ideal selves in a study with 60 participants. Across all three conditions (ESV, text-only, and mental imagination), we observed an increase in resilience, confidence, motivation, and goal commitment, and the ESV condition was perceived as uniquely engaging and personalized. We discuss the implications of designing generated self-voice systems as a personalized behavioral intervention for different scenarios.Published versio
Investigating lipid droplet-lysosome interactions during lysosomal stress
No full textLysosomes are essential degradative organelles that are responsible for the recycling of materials within the cell and overall maintenance of metabolic homeostasis. Lipid droplets (LDs) serve as dynamic reservoirs of neutral lipids within the cell. There has been emerging evidence indicating that LDs physically interact with lysosomes, a relationship that could have implications in lipid turnover, energy metabolism, and stress responses. While recent studies have characterised the existence of LD-lysosome contact sites, how these interactions change in the context of lysosomal damage has yet to be explored.
To investigate the changes in the LD population during lysosome damage, L-Leucyl-L-Leucine Methyl Ester (LLOMe) was employed to induce lysosomal membrane permeabilisation in HMC3 and U2OS cells. Neutral lipids within the cells were stained with 4,4-Difluoro-1,3,5,7,8-Pentamethyl-4-Bora-3a,4a-Diaza-s-Indacene (BODIPY) to facilitate LD identification, while 4’,6-Diamidino-2-Phenylindole (DAPI) staining was used to label the nucleus for single-cell segmentation. The LD distribution was then quantified at multiple time points following LLOMe treatment.
The experimental results demonstrated a consistent increase in LD content over time after LLOMe-induced lysosomal damage compared to untreated controls in HMC3-WT and U2OS-WT cells. Furthermore, live-cell imaging highlighted dynamic LD-lysosome interactions and revealed the potential internalisation of LDs in the lysosomal lumen during lysosomal damage. Thus, these preliminary results provide an insight into LDs functioning as lipid reservoirs that can aid in the repair of damaged lysosomal membranes.Bachelor's degre