DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
AI-powered chatbot with retrieval augmented generation (RAG)
No full textThrough the years, Artificial Intelligence (AI) Chatbots have undergone significant changes, shifting from their initial rule-based systems, to now, much more sophisticated Natural Language Processing (NLP) models, which are able to handle complicated queries and produce human-like speech. In spite of all these advancements, AI Chatbots are bottlenecked by limitations, for example, in their lack of consistency in output, and their instances of hallucinations.
Hence, my objective in this project is to fill in the gap between current AI Chatbot capabilities, and their potential of meeting the demands of niche domain expertise, specifically in a domain relevant to Nanyang Technological University (NTU) – an AI Retrieval Augmented Generation (RAG) Chatbot prototype designed to answer user queries about undergraduate programmes of the College of Computing and Data Science (CCDS) at NTU. The Chatbot helps parents and prospective students with their queries related to the niche domain of CCDS undergraduate admissions, illustrating the ability of the RAG-enhanced Chatbot to augment information retrieval, access that information through the relevant websites, as well as facilitate user engagement through the Chatbot experience.
The RAG Chatbot leverages web scraping to extract necessary information from the web, which it then processes with OpenAI’s embedding models to generate machine-readable representations. The Supabase vector database stores these embeddings, facilitating quick similarity searches. When a query is submitted, an independent question is formulated by the bot, which then extracts from Supabase the required context, and merges it with the query to generate an informed and precise output with the help of OpenAI’s language models. The architecture is built with Next.js and React on the frontend, while supported by the Supabase Schema on the backend.Bachelor's degre
Building multi-modal intelligence: dissecting model with monosemantic features and a unified evaluation approach
No full textWith the rapid advancement of Large Language Models (LLMs), the capabilities of Large Multi-Modal Models have been improving at an accelerated speed. However, despite these advancements, a substantial gap persists between these models and true human intelligence, indicating significant potential for further improvement. Creating a true intelligent agent requires more than just improving performance on static and out- dated benchmarks; it demands a holistic approach to refining the model’s capabilities throughout its entire development cycle. In this research project, we propose a full-cycle framework for model development, which includes the following key components: 1) Intersecting models to gain a deeper understanding of their intelligence levels through monosemantic features. 2) Establishing a standardized and unified evaluation method- ology to systematically assess model performance. 3) Constructing enhanced models based on insights derived from comprehensive analysis and rigorous evaluation.Bachelor's degre
ML and systems co-design for resource-efficient LLM inference serving
No full textThis project investigates the co-design of machine learning and systems-level techniques to enable resource-efficient inference serving of large language models (LLMs). We conduct a detailed analysis of the trade-offs between accuracy and performance (latency, memory, throughput) for a variety of model-level optimizations, including a modified version of Skeleton-of-Thought (SoT) prompting, prompt compression using LLMLingua, model quantization, and Key-Value Cache quantization. Through systematic experimentation and analytical modeling, we identify efficient configurations that minimize resource usage while maintaining output quality close to that of large models.
In addition to understanding the efficiency-accuracy trade-offs of the earlier mentioned model-level optimizations, the key contributions of this project also include the implementation of system components such as parallel expansion handlers, compression routers, and output routers—designed to integrate the evaluated optimizations into an adaptive, modular serving system. These components are built on top of the vLLM engine to leverage modern serving capabilities like continuous batching and paged attention.
Together, our findings and implementations lay the foundation for a complete LLM serving stack that can dynamically adapt to input characteristics and system load by leveraging various machine learning optimizations. The full integration of these components into a complete system remains ongoing work.Bachelor's degre
台湾男女形象的重构与男同困境——以电影《关于我和鬼变成家人的那件事》为例 = The reconstruction of gender images and predicament of gay men in Taiwan: a case study of “Marry My Dead Body”
No full text台湾在2019年正式通过了同性婚姻法案，成为了亚洲首个同性婚姻合法化的国家。这其中不乏台湾同志团体的积极推动和争取。当法律不再是限制在台同志收获爱情和幸福的限制后，在台同志是否就都取得了他们“圆满的结局”了呢？2022年上映的《关于我和鬼变成家人的那件事》以同性婚姻合法化后的台湾社会为背景设定揭露台湾男同性恋伴侣的困境。《关于我和鬼变成家人的那件事》通过诙谐的“直男警察与同志鬼冥婚”设定桥段刻画出不同性别形象，再以调查车祸事件的剧情去再现台湾社会对男同的偏见与接受度，以及男同之间的问题。笔者认为电影所企图展现的同性婚姻合法化后的台湾社会内涵和同志困境值得深究与思考。此外，电影里的男性与女性形象建构随着剧情的发展也有了不同的刻画，因此笔者将探讨电影其中性别形象的转变与意义。本文将以电影《关于我和鬼变成家人的那件事》为文本，通过分析文本探讨电影中所建构的性别形象与意义，再现了台湾男同面临的困境与其原因，以及题材的局限性。In 2019, Taiwan became the first Asian country to legalize same-sex marriage, a milestone largely driven by the efforts of Taiwan's LGBTQ+ community. However, after the legal barriers to love and happiness for same-sex couples were lifted, have Taiwanese LGBTQ+ individuals truly achieved their "happy ending"? The 2022 film “Marry My Dead Body” vividly portrays the challenges faced by Taiwanese gay couples in the post-legalization era. Through the humorous premise of a "gay ghost marriage" between a straight policeman and a deceased gay man, the film explores diverse gender identities and highlights the ongoing prejudice and acceptance struggles faced by gay men in Taiwan. As the film presents an insightful depiction of the invisible struggles and gender identity issues that persist after the legalization of same-sex marriage, it is necessary to further explore and analyse this topic in depth. Additionally, the evolving portrayal of male and female characters throughout the film offers a significant subject for analysis. Therefore, this essay analyses the film “Marry My Dead Body” to examine the construction and transformation of gender identities, the challenges faced by Taiwan LGBTQ+ community and their underlying causes, and the film’s thematic limitations.Bachelor's degre
Development of mobile applications for battery capacity detection in smart contact lens
No full textWith the development of wearable electronic devices, smart contact lenses (SCLs) have attracted increasing attention as a new tool for real-time health monitoring. One of the main challenges in advancing this technology is how to manage power effectively, especially when it comes to monitoring battery levels. Due to the small size and soft structure of SCLs, traditional battery monitoring methods are not suitable. Therefore, it is necessary to find new, simple, and non-invasive solutions.
This study introduces a software-based method for battery monitoring in SCLs that uses Prussian Blue (PB) as the electrochromic material. PB can change color during charging and discharging, which provides a useful signal to estimate the battery status. Based on this principle, A MATLAB-based image analysis system was developed to recognize the color of the lens and predict its battery capacity. The system includes image preprocessing, color feature extraction, and regression modeling using machine learning. A graphical user interface (GUI) was also created to visually reflect the data, which makes the application easier to use.
The method allows users to assess battery levels in real time without using complicated devices. It demonstrates a new way of combining electrochemical materials with digital tools and shows the potential for further development in wearable medical electronics.Master's degre
NGL... you should know it
No full text‘NGL… You Should Know It’ is a health communications campaign addressing the rising prevalence of Type 2 Diabetes in Singapore by promoting awareness and understanding of Nutri-Grade labels among 40 to 64-year-olds in Singapore.
Today, more than 400,000 people in Singapore live with diabetes, with this number projected to rise to 1 million by 2050. While diabetes remains a key priority in Singapore’s healthcare agenda, as seen in initiatives like the Nutri-Grade labels, there are still gaps in educating vulnerable audiences on how to use these labels effectively and understand the broader implications of diabetes.
By bridging the gap between label awareness and practical use through on-ground engagements, interactive experiences and social campaigns, we seek to empower individuals to take informed and practical steps toward making healthier beverage choices, ultimately reducing their risk of developing Type 2 Diabetes.Bachelor's degre
Silicon FET reservoir for dynamic edge vision
No full textAI is driving significant advancements in computer science and various
practical applications. However, its current reliance on centralized infrastruc-
tures and large-scale servers results in substantial power consumption and high
latency. To address these challenges, this article introduces a hybrid system that
integrates a physical reservoir with a software-based readout layer into a ResNet
neural network, specifically designed for dynamic edge vision applications.
We utilize a DVS-based edge processor, which helps offload certain tasks
from the centralized server and detect variations in object images. By leveraging
the ability of oxide vacancy defects to capture and release carriers, the H f O2
bulk exhibits memory-like behavior, converting pulse signals from the edge sen-
sor into analog values. These multidimensional values are then compiled into
frames, capturing both the temporal and spatial features of the original video.
For readout training and classification, we employ two types of neural net-
works: ResNet and ResNet-LSTM. Analyzing the output frames of the physi-
cal reservoir, the ResNet model achieves an average accuracy of 94.78% using
only a 2-second event stream, with a maximum accuracy of 96.00%. By incor-
porating the ResNet-LSTM model to analyze only the initial 3 seconds of the
event stream, using 0.5-second clips, we attain an average accuracy of 97.34%
and a maximum accuracy of 98.40%. These experiments are conducted on the
DVS128 dataset, classifying 10 gesture categories. Compared to existing models,
our approach achieves outstanding accuracy while maintaining relatively lower
computational weight.Master's degre
Beyond the ballot: persistence, people and possibilities in Singapore politics
No full textBeyond the Ballot is a journalism feature package that highlights the undying persistence of the people behind Singapore’s politics, including party volunteers and members of opposition parties who have been unable to win seats in Parliament in previous general elections. The package will feature the fortunes and misfortunes they have tackled throughout their journeys and why they continue fighting for their causes. In response to changing societal norms and expectations, it explores how women tackle bias and scrutiny as they fight for more representation in the political scene. And with the rise of social media, it also examines how parties have adapted to the growing relevance of social media in politics.Bachelor's degre
Characterization and analysis of semiconductor magnetic sensors: hall devices
No full textWith the development of technologies, sensors have been playing an essential role in the overall intelligent. In the sensor field, magnetic sensors are capable not only of detecting magnetic fields but also of measuring parameters such as pressure, position, angular velocity and current. In recent years, the magnetic sensor market has reached billions of dollars annually worldwide, with Hall effect sensors taking a significant share. This is primarily due to the fact that the detection range of Hall effect sensors aligns well with the magnetic fields generated by various materials. Moreover, Hall effect sensors are compatible with CMOS technology, offering a cost advantage over other magnetic sensors, as they can be integrated into a single chip alongside the signal conditioning circuitry and digital processing.
This project mainly involves characterization of advanced Hall sensors with the Future Rich CMOS team at GlobalFoundries Singapore. The focus is on studying the semiconductor physics related to the Hall effects based on detailed device characterization and analysis of test data obtained from the most advanced Hall sensor. Data collected from parametric analysers and other departments within GlobalFoundries are processed automatically using customized script, followed by further analysis. Effects of both bias and temperature are studied and analysis are presented. Moreover, Python scripts which are capable of automatically processing various raw data are developed, enabling hundreds of data being reformatted in seconds. Investigations based on obtained data could instruct engineers to develop more powerful devices. Continuous collaboration with other teams ensures alignment with industry standards and customer requirements, driving the advancement of semiconductor magnetic sensor technology and securing the leadership in manufacturing magnetic sensor in foundry industries.Master's degre
英汉情感心理动词 "爱" 和 Love 的对比研究 = A comparative study of emotional psych verb "爱" and love
No full text本文的主要研究为英汉的情感心理动词“爱”和Love。本研究旨在通过平行语料库分析和问卷调查，双向考察其在英语和汉语中的不同义项分布，分析其在中译英和英译中过程中所对应的义项是否一致，以及翻译中的词汇选择规律。研究结果显示“爱”对应率最高的是Love，而Love翻译为汉语时，对应率最高的也是“爱”。此外，Love在英语更常用于表达男女之间爱情的Love5 这个义项。通过问卷调查，本文揭示了双语能力、语义协调性及社会规范对翻译选择的影响，为理解“爱”和Love在双语环境中的翻译规律提供了新的视角，并为相关领域的研究和实践提供了参考。
This study focuses on the emotional psych verbs “爱” and “Love” in Chinese and English. Through a parallel corpus analysis and questionnaire survey, the research examines the distribution of their different meanings in both languages and analyses whether their corresponding definition remain consistent in Chinese-English and English-Chinese translation. The findings indicate that “爱” most frequently corresponds to “Love,” and vice versa. Additionally, “Love” is more commonly used in English to express romantic relationships (Love5). The questionnaire survey further reveals the influence of bilingual proficiency, semantic alignment, and social norms on translation choices. This study provides new insights into the translation patterns of “爱” and “Love” in a bilingual context, offering valuable references for future research and practice in related fields.Bachelor's degre