DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
LLM hallucinations in practical code generation: phenomena, mechanism, and mitigation
No full textCode generation aims to automatically generate code from input requirements, significantly enhancing development efficiency. Recent large language models (LLMs) based approaches have shown promising results and revolutionized code generation task. Despite the promising performance, LLMs often generate contents with hallucinations, especially for the code generation scenario requiring the handling of complex contextual dependencies in practical development process. Although previous study has analyzed hallucinations in LLM-powered code generation, the study is limited to standalone function generation. In this paper, we conduct an empirical study to study the phenomena, mechanism, and mitigation of LLM hallucinations within more practical and complex development contexts in repository-level generation scenario. First, we manually examine the code generation results from six mainstream LLMs to establish a hallucination taxonomy of LLM-generated code. Next, we elaborate on the phenomenon of hallucinations, analyze their distribution across different models. We then analyze causes of hallucinations and identify four potential factors contributing to hallucinations. Finally, we propose an RAG-based mitigation method, which demonstrates consistent effectiveness in all studied LLMs.Published versionThis work is supported by CCF-Huawei Populus Grove Fund CCF-HuaweiSE202403. This work is supported by the Guangdong Basic and Applied Basic Research Foundation (2023A1515012292)and CCF-Sangfor 'Yuanwang' Research Fund
Seeing through thick optical scattering media
No full textSeeing through highly scattering media remains a significant challenge in optical imaging, severely constraining applications in biomedical diagnostics, non-destructive testing, and remote sensing. Conventional optical methods encounter drastic resolution and contrast degradation due to multiple photon scattering, impeding direct optical path retrieval. This thesis explores Computational Time-of-Flight Diffuse Optical Tomography (ToF-DOT), an advanced imaging technique that combines time-resolved photon detection and computational reconstruction to overcome current limitations in optical tomography.
A comprehensive simulation framework based on MATLAB was developed to systematically evaluate reconstruction performance across a fixed optical thickness of 5 cm. Letter-shaped targets (``A'',``X'',``Y''), a triangle, and a complex facial silhouette were simulated within scattering media characterized by optical properties typical of biological tissues. Image reconstruction was achieved using gradient descent optimization enhanced with total variation (TV) and l1 regularization. Image quality was evaluated qualitatively through visual inspection, focusing on structural preservation, contour clarity, and edge sharpness. The results demonstrated successful reconstruction of primary object shapes, with notable degradation of finer structural details due to inherent photon diffusion.
This thesis demonstrates the effectiveness of total variation (TV) and l1 regularization in recovering spatial structures from time-resolved measurements under strong scattering conditions.Master's degre
Voices on non-residents: exploring Singaporean youths’ perspectives and experiences
No full textThis study explores factors that influence Singaporean youths’ attitudes towards non-
residents. This study uses a qualitative approach with semi-structured interviews to
analyse the impact of government policies, individual experiences, and social media on
these attitudes. Findings reveal that individual experiences have a more profound
influence than other factors. Building established relationships can be the key to
ensuring cohesion in the country’s community. Given the long-standing tension
between residents and non-residents, this study explores ways the government can
improve this relationship moving forward.Bachelor's degre
Learning based relocalization with UWB
No full textUltra-Wideband (UWB) technology is often celebrated for its accuracy in small-scale, indoor environments. However, even within these controlled environments, UWB faces significant limitations, such as susceptibility to non-Line-of-Sight conditions and multi- path interference, which can severely impair its efficacy. These challenges become even more pronounced in complex indoor spaces where traditional filtering methods struggle to maintain accuracy. First, anchors are deployed in the environment without knowing their actual position. Then, UWB observations are collected when the robot travels in the 6x6m indoor environment. At the same time, Position of the mobile robot is collected using a motion capture system to provide the ground-truth. This Final Year Project proposes an application of the Mamba model, a deep learning framework designed to en- hance UWB’s robustness against such environmental variabilities. By leveraging a unique dataset collected from a 6x6 meter indoor area, this study explores the potential of Mamba on the complexities of real-world signal interference. The Mamba model’s capability to learn from intricate patterns in UWB signal data promises a significant advancement in localization accuracy, potentially setting a new standard for UWB applications in densely structured indoor settings. Through this project, we aim to understand how well UWB can adapt to complex environments where traditional methods struggles, focusing on the application of deep learning techniques to refine and possibly extend the utility of UWB for precise localization.Bachelor's degre
Sensitivity profiles and their associations with emotional problems during a period of school transition: a latent profile analysis
No full textStudies have identified three subfactors of sensitivity as measured by the Highly Sensitive Child (HSC) scale: Aesthetic Sensitivity (AES), Ease of Excitation (EOE), and Low Sensory Threshold (LST). While studies have investigated how the subfactors are linked to different outcomes, there is a lack of research into how varying levels of these subfactors manifest in a sensitivity profile. Furthermore, there is a lack of studies that have investigated how sensitivity impacts the adjustment to a new school environment. This longitudinal study aims to fill in this gap by examining how children’s sensitivity profiles (n = 464; 52% female), measured by the HSC scale at primary school, predict emotional problems after transitioning to secondary school, and whether this relationship is mediated by school adjustment post-transition and prior emotional difficulties. Firstly, latent profile analysis uncovered three distinct sensitivity profiles: Moderate Reactivity (moderate AES, EOE, LST scores), Sensitive Stable (high AES, moderate EOE, and low LST scores), and High Reactivity (high AES, EOE, LST scores). ANOVA and path analysis revealed that compared to the Moderate Reactivity profile, only the High Reactivity profile had significantly greater direct associations with emotional problems post-transition. Compared to the Moderate Reactivity profile, both the Sensitive Stable and High Reactivity profiles predicted significantly better school adjustment, which in turn led to fewer emotional problems post- transition. However, compared to the Moderate Reactivity profile, only the High Reactivity profile predicted significantly greater emotional problems prior to school transition which subsequently explained greater emotional problems post-transition.Bachelor's degre
Learning electronic circuit via building a ChatGPT AI virtual assistant
No full textWith the progress in technology, different sectors have developed and progressed,
making full use of what technology has to offer. In the current day and age, all
sectors are now making use of Artificial Intelligence (AI) and Machine Learning
(ML) to automate and optimise their processes and products. The education sector
therefore cannot be left out. In this study, we propose to make use of Large Language
Models (LLM) to serve as “Virtual Assistants” that can serve and teach students at
any point in time and not just school hours.
This project utilises OpenAI’s “gpt-4” LLM which will serve to give us answers to
any input questions given. To further enhance a possible teacher-student relationship
with the Virtual Assistant, we will utilise voice detection and recognition software
like PicoVoice, which can not only detect specific words but also convert speech
audio files into text files which then can be fed into the OpenAI LLM. We also
produce the answer given not just in text form, but have it converted back into
natural, life-like sounding voices using Amazon Web Service (AWS) Polly service.
AWS Polly offers a multitude of different voices and accents that one can pick to
best suit the vibe of a teacher.
In addition to the Virtual Assistant, a webpage will be created using HyperText
Markup Language (HTML), Cascading Style Sheets (CSS) will be used to design the
webpage. PHP and Javascript will then be used to provide the frontend and backend
functions. The webpage will utilise a SQL database to store key information to
enable the webpage to give a more personal experience while operating the webpage.
The webpage will cater towards teaching electronic circuits by offering theory and
component lessons, a circulator and different practice questions for users to practice.
Overall, the webpage with the addition of the Virtual Assistant can allow users to
participate in lessons and practice with the aid of a reliable teacher at any place and
time.Bachelor's degre
A review on the impacts of artificial intelligence to the global supply chain (the port industry)
No full textAI is increasingly acknowledged as a pivotal element in all industries and the maritime sector is not excluded, particularly in revolutionizing the port industry. A significant transformation is
underway in the port industry, making it essential to assess the fusion socioeconomic and operational impacts within the field.
The report utilized two primary methodologies: surveys and interviews. A total of 50 survey participants from various roles across the supply chain provided insights, alongside three interviewees from maritime educational institutions who shared their experiences and perspectives.
The findings indicate that a significant portion of employees have a moderate understanding of AI, expressing concerns about its potential impact on their jobs. Many believe that AI implementation could negatively affect the workforce. However, there is a consensus that ports must adopt AI technologies to stay relevant in today’s advanced technological landscape, which
is vital for the workforce and will greatly influence the port industry. While the adoption of AI in container ports is widely acknowledged, its presence in multipurpose ports remains limited. AI has become integral to the fundamental operations of the port sector.
Three key discussions emerged from the findings. First, there is a challenge of technological uncertainty, as many employees lack familiarity and expertise in AI. Second, misconceptions about AI have been shaped by media portrayals, leading to a negative perception of its integration in the workplace. Third, the advent of the AI era has fundamentally transformed the structure of the port industry, creating a demand for new essential skills and future job roles, which may result in elimination in workers who fall into this vulnerable category.
Three key recommendations have been outlined. First, it is essential to establish a Human-AI Policy framework to guarantee the ethical development and application of AI technologies. Additionally, AI initiatives should prioritize societal benefits, emphasizing a human-in-the-loop or human-on-the-loop approach rather than excluding human involvement. Finally, a collaborative system involving Government, Education, and Industry should be developed to focus on creating job competency standards, designing an AI-adapted curriculum, and strengthening the teaching workforce in this field to cultivate the skilled professionals necessary for the advancement of AI.
The core principle of "humanistic" collaborative development should guide the interaction between AI and human intelligence.Bachelor's degre
Label efficient learning for video grounding
No full textGiven a natural language query, the task of video grounding aims to identify the visual content in an untrimmed video described by the language query. It is one of the most fundamental tasks in video understanding and has a wide range of real-world applications, such as robotic manipulation and video surveillance analysis. In recent years, video grounding has been remarkably advanced by the deep learning techniques and the availability of massive data. However, existing approaches depend heavily on manual video annotation, which is labor-intensive and unscalable, limiting their real-world applicability. This thesis addresses label-efficient learning for video grounding, aiming to reduce manual annotation costs while maintaining competitive performance.
The first work of this thesis investigates the limitation of an established labelefficient paradigm, specifically the weakly-supervised setting for video grounding. We find that current approaches in this setting often compromise performance due to insufficient supervision. To address this issue, we focus on utilizing complementary information from multi-modal videos such as RGB frames and optical flows, which naturally introduces richer supervision in weakly-supervised contexts. We propose a multi-modal distillation algorithm to exploit the multi-modal knowledge as supervision for model training. Experiments on two large-scale datasets demonstrate that our method significantly improves the performance of weakly-supervised learning.
Different from the first work that focuses on auxiliary supervision from the visual side of multimodal videos, the second work addresses it on the language front. We investigate the challenge of limited input distribution in language queries under weak supervision, where the narrow writing styles of human-annotated queries hinder the model’s ability to generalize to real-world scenarios that involve varied vocabularies and sentence structures. To overcome these challenges, we propose an omnipotent distillation algorithm with large language models (LLMs). The distribution of the input sample is enriched to obtain diverse multi-view versions while a consistency loss then comes to regularize the consistency of their results for distillation. Our experiments show substantial performance improvements when adapting to diverse language queries.
The aforementioned two works improve weakly-supervised video grounding by introducing auxiliary supervisory signals. However, annotating weak labels, such as sentence queries, remains necessary in these two works. To further relieve the annotation costs, the third work explores a novel paradigm by pretraining the temporal video grounding model on unlabeled videos. To support this, we introduce Vid-Group, a large-scale dataset collected with minimal human intervention for video grounding pretraining. To tackle the issues of error-prone pseudo training samples, we propose the ReCorrect algorithm. ReCorrect incorporates semantics-guided refinement to clean and adjust pseudo labels and exploits memory consensus correction to calibrate temporal boundaries based on consensus within a memory bank. Comprehensive experiments demonstrate ReCorrect’s strong generalization abilities across multiple downstream settings. For instance, zero-shot ReCorrect achieves over 75% and 80% of the best fully-supervised performance on two benchmarks, while unsupervised ReCorrect reaches about 85% on both.
Training data can be inaccessible in many real-world applications, such as due to privacy concerns. To address this issue, the final work makes a first exploration of video grounding in a training-free manner, eliminating the necessity of any training videos or annotations. A significant challenge in video grounding lies in the demand for joint spatio-temporal reasoning. To achieve this, we propose an Expectation Maximization Multimodal Modulation (E3M) algorithm, which comprises a context-based visual modulation, a prototype-based textual modulation, and an EM optimization in an iterative paradigm. Our training-free E3M approach outperforms several state-of-the-art methods with stronger supervision across three large-scale benchmarks.Doctor of Philosoph
Vision-based 3D human action assessment for stroke rehabilitation
No full textRecovering from stroke requires consistent physical assessments to track patient
improvement. Traditionally, these assessments are manual, subjective, and timeintensive,
which lead to inconsistency and inefficiency.
This project, developed with Tan Tock Seng Hospital, presents
an automated scoring system that uses 3D motion-tracking
and analysis — combining human pose reconstruction with
object interaction tracking — to evaluate stroke recovery
more objectively and efficiently
Who am I: the journey of rediscovering oneself
No full textThe motif of the journey has been, and remains prevalent in literary works. It appeals to our “desire for purpose” and the “triumph of good over evil” as we identify with the characters embarking on their adventures (Saluja et al. 1084). The journey allows readers to “explore the unknown” in imagined worlds (1083). It may take place in an unfamiliar environment and involve characters encountering obstacles before they reach their destination. A similar narrative is travel writing, where the writer discovers a place, experiences its physical space, and reflects on the details of the journey (qtd. in Galang-Pereña 102). Writers may interpret the journey based on their own past and beliefs, revealing new meanings. Travel writing includes “testimony” and “autobiography”, which are personal (103). They also allow writers to understand themselves after they are introduced to new places and people (103). This implies a link between the physical journey and selfhood, which this thesis shall discuss.Bachelor's degre