DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
Comparative insights into major TCM databases and their biomedical integration
No full textFollowing the discovery of artemisinin by Professor Tu Youyou and team for novel treatment of malaria, more studies have been conducted to identify constituents and explore pharmacological activities of herbal medicines. Over the past 2 decades, there has been increased focus on the creation and development of TCM-centric databases from various multi-omics studies, improving comprehensiveness and standardisation of TCM-related information. While these databases empower researchers to better systematically understand herb/formula-ingredient-body interaction, there are many different databases, each using its own unique data acquisition, extraction and processing methodology, resulting in a mix of data, not standardised across databases.
Hence, it is a necessity to review these databases, providing insight into the state of the field of TCM databases through analysing genealogy, functionality and sources used by the 12 databases compared. Doing so allows for future improved data harmonisation and integration, better utilising available information, and provides an exploratory map for interested researchers to navigate the world of computational TCM.
Through this study, I aim to highlight and compare available TCM databases through machine learning algorithms, providing an overview and analytical insights into the state of TCM databases. Thereby, improving overall understanding of TCM by the larger scientific community.Bachelor's degre
Metagenomic survey and identification of potential saprophytic wood decay fungi in Nanyang Technological University
No full textWith the increasing popularity of timber as a building material, there is a need to understand
the local diversity of saprophytic wood-decay fungi in Singapore, which has not been well
studied. This study focuses on fungal populations in decaying wood structures in Nanyang
Technological University (NTU), with the objective to identify the main players in wood
decay.
Metagenomic analysis of the samples collected yielded a preliminary database of fungi
populations potentially involved in wood decay, with quantitative and qualitative
information. 14 genera of known wood-decay fungi were found in decaying wood planks and
fruiting bodies in NTU. These fungi generally differed from wood-decay fungi found in
Europe, notably the absence of S. lacrymans. The fungal composition of GAIA, made of
treated coniferous wood, was different from outdoor wood structures within NTU, made of
untreated tropical wood. Metagenomic analysis also revealed unidentified species that
potentially play a role in wood decay.
In vitro wood decay analysis was conducted for Spruce, Rubber, Angsana and Balau wood
using two isolated strains of saprophytic wood-decay fungi (P.maximus, S.commune). Decay
was measured after 2 weeks, with prominent loss of dry weight in Angsana. More conclusive
results may be found after 2 months.Bachelor's degre
Analysis of transboundary haze on Singapore: a computational fluid dynamics approach
No full textThis study investigates the behavior of the Transboundary Haze across
Singapore, a Computational Fluid Dynamic study of the airflow. The erratic
behavior of the haze, how it spreads across an urban environment of large
structures, the circulation of it within the area and what causes it to stagnate.
The report explains the literature review regarding airflow behavior and its
difficulties posed with clearly defined limitations on the subject. The report
further presents the methodology used to capture these behaviors utilizing
the available knowledge and technologies.
The behavior of the Transboundary Haze was simulated using CFD over a
model of Marina Bay Sands and its surroundings to be used as analysis over
Singapore’s urban environment. The identified circulation, dispersion and
retention of airflow and its impacts were noted down in the study.
This study concludes with recommendations for Singapore’s current
circumstances and insights to ways the study could be used for other
applications.Bachelor's degre
Prettismart: visual interpretation of smart contracts via simulation
No full textSmart contracts are the fundamental components of blockchain technology. They are programs to determine cryptocurrency
transactions, and are irreversible once deployed, making it crucial for cryptocurrency investors to understand the cryptocurrency
transaction behaviors of smart contracts comprehensively. However, it is a challenging (if not impossible) task for investors, as they donot necessarily have a programming background to check the complex source code. Even for investors with certain programmingskills, inferring all the potential behaviors from the code alone is still difficult since the actual behaviors can be different when differentinvestors are involved. To address this challenge, we propose PrettiSmart, a novel visualization approach via execution simulation toachieve intuitive and reliable visual interpretation of smart contracts. Specificall, we develop a simulator to comprehensively capturemost of the possible real-world smart contract behaviors, involving multiple investors and various smart contract functions. Then, wepresent PrettiSmart to intuitively visualize the simulation results of a smart contract, which consists of two modules: The SimulationOverview Module is a barcode-based design, providing a visual summary for each simulation, and the Simulation Detail Module is anaugmented sequential design to display the cryptocurrency transaction details in each simulation, such as function call sequences,cryptocurrency flws, and state variable changes. It can allow investors to intuitively inspect and understand how a smart contract willwork. We evaluate PrettiSmart through two case studies and in-depth user interviews with 12 investors. The results demonstrate theeffectiveness and usability of PrettiSmart in facilitating an easy interpretation of smart contracts.Ministry of Education (MOE)Submitted/Accepted versionThis project is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Proposal ID: T2EP20222-0049)
Exploiting social relationships for robust recommendation
No full textSocial recommendation (SR) addresses data sparsity and cold-start challenges by leveraging users’ social connections, assuming that socially linked users tend to share similar preferences. Early SR models were primarily based on matrix factorization and more recent methods have adopted neural structures like graph neural networks (GNNs) to model higher-order social dependencies, and self-supervised learning (SSL) to exploit auxiliary signals in user-item or social graphs.
Despite these developments, real-world social networks are often misaligned with users’ true preferences, as well as being noisy and sparse, which undermines the reliability of social signals. While most SR methods treat the social graph as a fixed input and focus on architectural enhancements, this thesis approaches SR from a robustness perspective, where robustness is defined as the model’s ability to maintain reliable performance in the presence of imperfections in the social structure. Instead of passively incorporating social links, we explore how adapting and refining the social graph itself can enhance the reliability and effectiveness of social recommendation.
These approaches have significantly improved accuracy, but most still treat the social network as a fixed input and rarely question its quality or consistency. In contrast, this thesis investigates SR from a robustness perspective, aiming to adapt and refine social structures themselves to improve the reliability and effectiveness of social recommendation.
In social recommendation, user behaviors can be categorized into item consumption and social connection, forming two interrelated networks: the user-item interaction and the social network. These behaviors stem from distinct motives, each shaping users' preferences differently. However, existing models typically learn holistic embeddings that entangle these diverse motives, leading to less robust representations that are more susceptible to noise and irrelevant signals. Moreover, most studies address item consumption and social link prediction separately while overlooking their mutual influence. To address this, we propose DMJP, which disentangles the motives behind item consumptions and social connections for mutually enhanced joint prediction. Specifically, we first learn multi-faceted user representations in both networks to enhance robust preference inference. Then, we discriminate mutual influence between them at the facet-to-facet level and exploit it asymmetrically to refine user representations. Empirical studies on three public datasets showcase the superiority of DMJP over state-of-the-arts (SOTAs) on both tasks.
While disentangling motives behind users' social connections can facilitate robust preference inference, the complex user behaviors can introduce noisy connections that misrepresent preferences and degrade recommendation performance. Thus, noisy edges should be removed to prevent them from compromising the model robustness. However, without noise labels, existing denoising methods rely on supervised signals from noisy social networks, making the relevance of retained connections to user preference modeling uncertain. To address this, we propose SSD-ICGA, a self-supervised denoising approach for robust SR. Specifically, we introduce an independent cascade-based graph augmentation that simulates cascading influence spread in the social network, seamlessly amplifying the impact of Opinion Leaders (OLs) and high-similarity neighbors while weakening that of low-similarity neighbors through a hierarchical contrastive task. Such augmentation is further integrated with an information bottleneck-based contrastive task for effective self-supervision towards a denoised social graph. Experiments conducted on two public datasets demonstrate that SSD-ICGA outperforms SOTAs while also exhibiting higher robustness to different extents of social noise.
In addition to the noisy social connections, the effectiveness of SR is also constrained by the limited number of social connections for most users, which hampers the system's ability to fully leverage social influence. Therefore, it is essential to perform social network refinement by removing noisy connections and adding meaningful ones for robust SRs. Inspired by the denoising capability of generative diffusion models, we propose a model-agnostic social network refinement framework with diffusion models for robust social recommendation (ARD-SR). Specifically, we introduce a position-specific noise schedule in the forward process and guide the reverse process with neighborhood preferences, enabling robust and personalized social network refinement. We further design a curriculum-based training mechanism to gradually introduce more challenging samples, effectively boosting learning efficiency. Experiments on three datasets show that ARD-SR significantly improves SR performance across multiple SR backbones.
To sum up, in this dissertation, we propose a series of novel recommendation approaches by exploiting the social network for robust recommendations. Extensive experiments on real-world datasets show the effectiveness of our proposed methods.Doctor of Philosoph
AI attacks AI: recovering neural network architecture from NVDLA using AI-assisted side channel attack
No full textDuring the last decade, there has been a stunning progress in the domain of Artificial Intelligence (AI) aided by highly trained Machine Learning (ML) models. Such models are valuable Intellectual Property (IP) and, therefore, have been subjected to various model recovery attacks. In this work, we study the vulnerabilities of commercial, open-source accelerator NVDLA and present the first successful model recovery attack. For this purpose, we used power and timing information from the side-channel leakage of convolutional neural networks (CNN) models to train CNN-based attack models. Utilizing these attack models, we demonstrate that even with a highly pipelined architecture, multiple parallel execution in the accelerator along with Linux OS running tasks in the background, recovery of number of layers, kernel sizes, output neurons and distinguishing different layers, is possible with very high accuracy. This is also the first work to show the impact of differences in hyperparameters on the power traces.
Our solution is fully automated, AI-based, and portable to other hardware neural networks, thus presenting a greater threat toward IP protection. Using LeNet as the target victim model, we demonstrate an accuracy of more than 95% in recovering various parameters. This study presents a serious practical threat, in the form of side-channel attack, towards complex commercial architectures. Furthermore, we show that AI-guided attack significantly boosts the attacker capability.Submitted/Accepted versio
Easibility study of piezoelectric energy generation for Singapore land transport industry
No full textThis project aims to explore the potential of piezoelectric energy harvesters for generating and storing electricity in the Singapore land transport industry. By storing charge into a battery, these devices can serve as an alternate power supply for loads, thereby reducing the demand from the grid. The report would assess the feasibility of piezoelectric energy harvesters' performance in various environmental conditions specifically in Mass Rapid Transit (MRT) stations, optimizing efficiency, and researching on scalable production methods to reduce costs. Additionally, the project will focus on designing prototypes tailored to specific applications, which aim to maximize their energy outputs and making them compatible with the existing infrastructure, as well as performing cost comparison with other forms of energy harvesting systems and products currently available commercially.
Through these efforts, the project seeks to unlock the full potential of piezoelectric energy generation for sustainable and resilient power solutions.Bachelor's degre
Cellular glycocalyx affects nanoparticle access to cell membranes and uptake
No full textUnderstanding nanoparticle interactions with cells is fundamental to designing them for medical applications. Nanoparticles must interface with the cell surface to be bound and taken up. The glycocalyx is a carbohydrate layer coating the cell surface, rendering it negatively charged. Many researchers have noted that the glycocalyx affects nanoparticle uptake, but the mechanism remains unknown, Here, we investigate the interaction between the glycocalyx and nanoparticles at the cell surface in different cell types. The glycocalyx reduced the interactions between the nanoparticles and cells, thereby reducing cellular access, binding, and uptake. The magnitude of the effect is dependent on the nanoparticle charge. Fine-tuning the charge of nanoparticles can enhance the specificity of nanoparticle targeting. Understanding the role of the glycocalyx in nano-bio interactions will allow researchers to control the interactions of nanoparticles with the cell surface.Submitted/Accepted versionThis work was supported by the Canadian Institute of Health Research (FDN159932 and MOP-1301431), NanoMedicines Innovation Network (2019-T3-01), and the Canadian Research Chairs Program (950-223824). The author thanks the Ontario Graduate Scholarship (B.B. and M.G.G.M.), and the Cecil Yip Award (B.B.)
Navigating the future: anticipating geopolitical, financial, and global supply chain disruptions in the shipping industry (financial)
No full textThe shipping industry is facing unprecedented financial and operational challenges due to the European Union’s Emissions Trading System (EU ETS) and FuelEU Maritime regulations, which impose stringent carbon pricing and greenhouse gas intensity reduction targets. These policies significantly increase compliance costs for shipping companies, particularly those relying on traditional fossil fuels. This study assesses the financial implications of these regulations, evaluating the impact of various fuel choices including Very Low Sulphur Fuel Oil (VLSFO), Liquified Natural Gas (LNG), methanol, and ammonia on operational costs and regulatory compliance. A case study demonstrates how compliance costs escalate
over time, making conventional fuels financially unsustainable, while alternative fuels present varying
degrees of cost-effectiveness and infrastructure challenges.
To mitigate these financial burdens, the study explores strategic responses adopted by shipping
companies, such as fleet renewal, fuel diversification, operational efficiency improvements, and carbon credit trading. While short-term strategies focus on voyage optimization and carbon cost hedging, medium-term approaches involve fuel transitions and long-term supply agreements. The long-term shift towards ammonia and zero-carbon fuels is identified as the most viable compliance pathway. However, the study also highlights critical flaws in the current global decarbonization framework, including carbon market volatility, fuel availability disparities, regulatory inconsistencies, and carbon leakage risks.
This research concludes that a phased approach to decarbonization is essential for financial sustainability.
Furthermore, policy adjustments, including price stabilization mechanisms, infrastructure investments, and regulatory harmonization, are necessary to ensure a more predictable and equitable transition for the maritime industry. The findings provide valuable insights for shipping companies and policymakers navigating the evolving regulatory landscape.Bachelor's degre
Frontier detection using a machine learning approach
No full textThis dissertation focuses on improving multi-robot exploration using machine learning in unknown environments. Multi-robot exploration methodologies often require high computational complexity, which limits their use in real-time tasks. To solve this problem, we propose a reinforcement learning-based approach that combines Multi-Agent Proximal Policy Optimization (MAPPO) with a graph transformer. This design helps robots make better decisions and cooperate more effectively, resulting in improved performance in terms of exploration path length, reduced standard deviation of explored areas between the two robots, and lower exploration overlap. These improvements are especially significant in complex environments, compared to the baseline model which only employs the MAPPO algorithm. The framework is trained and tested in different simulated environments. Results show that the proposed model improves exploration efficiency in most cases. Compared with baseline model, it achieves a 32.2% shorter path length, lowering the overlap rate by 53.59%, and decreasing the standard deviation by 28.37% in room1. The model works especially well in complex environments with many corners. In conclusion, this work shows that using reinforcement learning with reasonable design of reward function and graph attention can help robots explore unknown areas more efficiently. Some problems still remain in simple environments, and future work will focus on improving the model and testing it on real robots.Master's degre