DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
Eh is it the same or not?: An analysis of difference in Singlish use across generations
No full textThis thesis seeks to use the Corpus of Singapore English Messages (CoSEM) to find out if there are significant differences in the use of Singlish across different age groups, namely 17-19-year-olds (Y group) and 33-67-year-olds (O group), with a total of 878,579 words between them. This builds on previous studies that used CoSEM for other linguistic features, while this thesis seeks to provide an overview of the comparison of Singlish features between the age groups. The results have been distilled into 5 categories, namely lexical borrowings, particles, shortenings, laughter and the use of media. The analysis revealed the Y group uses more modifications to Singlish, like shortenings to express themselves, while at the same time being more emotive in their use of laughter in their messages. The O group uses more lexical borrowings and surprisingly send more media, which could be a result of a larger linguistic repertoire as compared to the Y group but shows a difference in use for some which could be argued to be incorporated into the Singlish lexicon. Therefore, significant differences attributable to age differentiation in the use of Singlish can possibly demonstrate that Singlish has indeed changed over time.Bachelor's degre
Development of a wireless sensor for vibration measurement I
No full textThis report presents the Final Year Project on the development and design refinement of the Railway Sensor used in train vibration monitoring systems. This project addresses two key challenges: The potential interference of static electricity on railway tracks, which in turn would potentially affect the vibration measurements, and the ingress of solid impurities like dirt and moisture into the sensor enclosure. To address the potential interference of static electricity on railway tracks, experimental testing was done in the Mechanics of Machinery Laboratory using a railway track fixture to understand the influence of static electricity on the vibrations measured by the Railway Sensor. Separate experiment testing was also conducted in the Protective Engineering Lab using the Large Shaker to test the effects of different clamping scenarios on the railway track. This would help to better understand the behaviour of the clamps used during maintenance of the railway tracks, in the events of railway cracks. To tackle the potential ingress of impurities into the sensor enclosure, several design iterations were proposed to enhance the Railway Sensor’s functionality, focusing on ease of installation, level of protection against impurities, and durability. The final design, featuring a modular body made from 3D Printed Polytetrafluoroethylene (PTFE) material, integrated sealing features, and improved ergonomics for easier deployment on the railway tracks. Performance tests confirmed the effectiveness of ingress protection, without any loss of electrical performance. These optimised solutions would help to improve both the accuracy of data obtained, and the operational longevity of the Railway Sensor, contributing in more reliable railway condition monitoring and maintenance planning.Bachelor's degre
Environment data processing for data centre (1)
No full textData centres are the backbone of our digital society, powering cloud computing, artificial intelligence, and countless online services. However, they are also significant
energy consumers, with cooling systems representing a large fraction of operational
costs and environmental impact. In tropical climates, where ambient temperatures and
humidity levels are elevated, efficient thermal management is critical to ensure both
energy efficiency and equipment reliability.
This study explores the energy consumption modelling of an air free-cooled data center
using a controlled testbed, referred to as TDC 1.0. By leveraging Physics-Informed
Machine Learning (PIML), the project integrates established physical laws into various
machine learning models to enhance prediction accuracy and generalization. The
testbed is instrumented with sensors that capture key parameters such as temperature,
relative humidity, airflow rates, and power usage across different controlled temperature
setpoints.
Various data preprocessing techniques are applied to address challenges like timestamp
misalignment and outliers. Multiple machine learning approaches, including polynomial regression, multilayer perceptron (MLP), Random Forest, and XGBoost are
evaluated in both their baseline and physics-informed forms across varying sampling
intervals. The physics-informed models incorporate domain-specific loss functions that
blend empirical errors with deviations from theoretical physical predictions, thereby
guiding the learning process towards physically consistent outcomes.
The findings show that incorporating physical laws into the modeling process not
only improves robustness against noisy sensor data but also yields more physically consistent predictions of cooling energy consumption. Notably, the physics-informed
MLP consistently outperformed its baseline counterpart, with improvements reaching
up to 14% in R² at shorter sampling intervals, making it the most effective model for this
application. Particularly in environments characterized by high-frequency noisy sensor
data, the PIML approach provides a more reliable estimation of cooling demands,
enabling data centre operators to optimize energy usage and reduce costs. Future
work will focus on extending the physics-based framework by incorporating additional
physical laws and exploring advanced data augmentation techniques to further refine
model performance under diverse operating conditions.Bachelor's degre
IOT for plant condition monitoring
No full textMonitor plant health by analyzing daily variation in leaf greenness.
Use RGB image data to quantify chlorophyll related green intensity.
Identify Stress/ recovery trends through time series tracking.
Enable real time, scalable monitoring for use in precision agriculture
How interfaith dialogue upholds the fabric of multiculturalism in the face of disenchantment and hatred
No full textInterfaith dialogue cultivates trust, counteracts fragmentation, and nurtures shared values in diverse societies. It extends beyond mere discussion, requiring meaningful encounters with people in our daily lives. By embracing pluralism with humility, spirituality becomes a force for justice, bridges polarisation, and builds compassionate public spaces amidst rising cynicism and disenchantment.Published versio
Toward trustworthy self-driving systems: evaluation, decision, and cooperative control
No full textSelf-driving technology holds immense potential to revolutionize future mobility,
optimizing the utilization of traffic facilities while minimizing human effort. Despite
the remarkable progress achieved through machine learning, existing autonomous
driving algorithms face significant challenges in attaining human-like, trustworthy,
and safe behaviors within complex, dynamic, and uncertain open-world environments.
This thesis adopts a game-theoretic approach to address these challenges
and advance toward trustworthy self-driving systems, as in many existing safetycritical
systems.
This thesis is structured as a three-player game involving the evaluation, the decision,
and the collaboration. 1) Evaluation: The evaluation’s role is to identify vulnerabilities
and flaws within the existing self-driving algorithm. 2) Decision: The
decision focuses on enhancing the algorithm’s performance, continuously evolving
in response to the evaluation. 3) Collaborative Control: The collaboration’s objective
is to work alongside the reasoning to overcome challenges that could jeopardize
the reliability and safety of the self-driving system.
Regarding the evaluation, two types of adversaries are proposed: adversarial scene
generation and online monitoring. Unlike conventional scene generation algorithms,
adversarial scene generation aims to behave in a human-like manner while simultaneously
introducing ambiguity for the self-driving algorithm. The interaction
among traffic users is modeled as a Bayesian Game, targeting scenes with multiple
Nash equilibriums to proactively challenge the subject under test. Through
comparisons with existing scene generation algorithms and a Turing test, the proposed
method is demonstrated to be both human-like and effective in generating
diverse adversarial behaviors by altering a few key parameters. Furthermore, the
adversarial scene generation algorithm is enhanced to automatically generate potentially
harmful, long-tail, and ambiguous scenes. Comparative analysis and ablation
studies validate the algorithm’s effectiveness in generating highly critical traffic
scenes and identifying limitations within self-driving algorithms, regardless of the
framework. Additionally, recognizing the constraints of available datasets and the
generalization limitations of current learning-based algorithms, a planner-agnostic
monitor is proposed to supervise the self-driving algorithm online and issue alerts
to highlight deficiencies. This monitor is trained adversarially, employing a generator
to both challenge the monitor with out-of-distribution instances and evaluate
its own limitations. The monitor is trained on the Argoverse 2 dataset and validated
using HighD, NGSIM, RounD, and NuScenes. Results demonstrate that the
proposed monitor effectively supervises the self-driving algorithm online, regardless
of the self-driving framework, while achieving a reasonable trade-off between false
positives and false negatives compared to the two existing methods.
For the decision, an online learning approach is proposed for the prediction module
to dynamically adapt to the behavior of traffic participants in real time using
partial observations. This method aims to bridge the gap between static datasets
and the complexities of real-world scenarios. Simultaneously, the proposed algorithm
encourages surrounding road users to adapt to the ego vehicle’s decisions by
introducing an entropy reward. Inspired by the second law of thermodynamics,
this reward promotes decisions that reduce environmental chaos, ultimately fostering
unanimous predictions. Through human-in-the-loop testing, the algorithm has
been shown to effectively manage dynamic environments while ensuring the ego vehicle’s
safety. The algorithm is trained exclusively on the Argoverse 2 dataset and
rigorously tested across four additional datasets via ablation studies. The results
demonstrate the algorithm’s potential to address the disparity between training
data and real-world applications, reinforcing its adaptability and robustness.
Building on the problems identified by the evaluation, two major solutions are proposed
to enhance collaboration between humans and autonomous systems. The
first solution is a collaborative framework that employs indirect shared control as
a medium for human intervention. This framework targets multimodality—and
consequently ambiguity—in prediction and decision-making. Predictions and decisions
are modeled using a generative adversarial neural network to capture the
inherent variability. The control authority is determined by the diversity in predictions,
while haptic force feedback is adapted online based on the distribution of
decisions. Results from human-in-the-loop experiments suggest that this method
has significant potential to mitigate human-machine conflicts and assist self-driving
cars in making informed decisions in highly multimodal traffic scenarios, leveraging
high-level human intelligence. The second solution addresses out-of-distribution,
personalized driving challenges and reduces the reliance on driving expertise. A
novel gesture-projection-based human-machine interface is introduced, designed
to incorporate high-level human intelligence in a natural and explainable manner
that aligns with social norms. This approach enables intuitive guidance for the
vehicle. Human-in-the-loop experiments using EEG sensors to monitor mental
workload and EMG sensors to evaluate physical workload demonstrate that the
proposed method significantly reduces physical effort compared to conventional
manual driving. Importantly, this is achieved without compromising safety or
increasing mental workload, highlighting its potential to enhance human-vehicle
collaboration in complex environments.
The proposed game-theoretic framework for closed-loop and continuous improvement
of self-driving algorithms represents a promising step toward developing autonomous
and trustworthy systems. Within this framework, both natural and
adversarial scene generation algorithms are introduced, taking into account compound
error propagation and human-likeness evaluation. In parallel, a novel online
monitoring module is designed to assess system reliability in real time. To address
ambiguity in prediction, online learning and shared control strategies are
proposed. Furthermore, a gesture-based human-machine interface is developed to
handle out-of-distribution scenarios and enable personalized interaction. By tackling
these critical challenges, the proposed framework holds strong potential to
enhance public trust and support the widespread adoption of future autonomous
mobility solutions.Doctor of Philosoph
Iridium-catalyzed ester-directed oxidative coupling of aromatic esters and alkenes
No full textThe directing group-assisted, transition metal-catalyzed oxidative coupling of arenes with alkenes is a powerful tool for synthesizing functionalized alkenes. In this work, we report the iridium catalyzed oxidative coupling of aromatic esters with alkenes.Ministry of Education (MOE)Nanyang Technological UniversitySubmitted/Accepted versionThis work was supported by Nanyang Technological University and the Ministry of Education Singapore (Research Grant No. RG89/23)
Eliminating backdoors in neural code models for secure code understanding
No full textNeural code models (NCMs) have been widely used to address various code understanding tasks, such as defect detection. However, numerous recent studies reveal that such models are vulnerable to backdoor attacks. Backdoored NCMs function normally on normal/clean code snippets, but exhibit adversary-expected behavior on poisoned code snippets injected with the adversary-crafted trigger. It poses a significant security threat. For example, a backdoored defect detection model may misclassify user-submitted defective code as non-defective. If this insecure code is then integrated into critical systems, like autonomous driving systems, it could jeopardize life safety. Therefore, there is an urgent need for effective techniques to detect and eliminate backdoors stealthily implanted in NCMs. To address this issue, in this paper, we innovatively propose a backdoor elimination technique for secure code understanding, called EliBadCode. EliBadCode eliminates backdoors in NCMs by inverting/reverse-engineering and unlearning backdoor triggers. Specifically, EliBadCode first filters the model vocabulary for trigger tokens based on the naming conventions of specific programming languages to reduce the trigger search space and cost. Then, EliBadCode introduces a sample-specific trigger position identification method, which can reduce the interference of non-backdoor (adversarial) perturbations for subsequent trigger inversion, thereby producing effective inverted backdoor triggers efficiently. Backdoor triggers can be viewed as backdoor (adversarial) perturbations. Subsequently, EliBadCode employs a Greedy Coordinate Gradient algorithm to optimize the inverted trigger and designs a trigger anchoring method to purify the inverted trigger. Finally, EliBadCode eliminates backdoors through model unlearning. We evaluate the effectiveness of in eliminating backdoors implanted in multiple NCMs used for three safety-critical code understanding tasks. The results demonstrate that EliBadCode can effectively eliminate backdoors while having minimal adverse effects on the normal functionality of the model. For instance, on defect detection tasks, EliBadCode substantially decreases the average Attack Success Rate (ASR) of the advanced backdoor attack from 99.76% to 2.64%, significantly outperforming the three baselines. The clean model produced by EliBadCode exhibits an average decrease in defect prediction accuracy of only 0.01% (the same as the baseline).AI SingaporeNational Research Foundation (NRF)Published versionThis research is supported by the National Research Foundation, Singapore, and DSO National Laboratories under the AI Singapore Programme (AISG Award No: AISG2-GC-2023-008), the National Natural Science Foundation of China (61932012, 62372228, U24A20337), the Fundamental Research Funds for the Central Universities (14380029), the Open Project of State Key Laboratory for Novel Software Technology at Nanjing University (Grant No. KFKT2024B21), and the Science, Technology and Innovation Commission of Shenzhen Municipality (CJGJZD20200617103001003, 2021Szvup057)
Was Iran denuclearised by the Israeli and US air strikes?
No full textIsrael and the US attacked Iran’s nuclear facilities recently. The military strikes might have taken out or damaged Iran’s uranium enrichment stockpile and facilities, but its nuclear ambition can only be shaped through effective nuclear diplomacy.Published versio
Reasserting ASEAN's relevance: the road ahead
No full textThe upending of the American-led post-Cold War international order should prompt ASEAN to adopt practical measures to reassert its “centrality”. These measures should begin with improved management of key regional security challenges, both traditional and non-traditional, amid global geopolitical uncertainty.Published versio