Research Experiences

Neuroscience Student Researcher

University of Cambridge, Department of Engineering, Oct 2022 - Jun 2023

Joined the Neural Dynamics and Control (NDC) group directed by Dr Guillaume Hennequin.
Our task is to implement a webcam-based eye tracker to track the user’s eye movements on screen in psychophysical experiments to better understand human thinking when solving challenges.
Using ideas from human physiology, we implemented an online process that dynamically tracks environment variables such as the user’s viewing distance from the screen with one initial simple calibration task.
Based on WebGazer, we map from the user’s eye images in the camera stream (from TensorFlow Facemesh) to eyeball rotation angles using Gaussian Processes regression with a novel proposed kernel.
Our implementation shows a 22% reduction in error in prediction of gaze location of user compared to the original WebGazer. View My Report. Check out the project at GitHub Repo.
Supervisor: Dr Guillaume Hennequin

University College London, Department of Computer Science, Dec 2021 - Oct 2022

Joined the Web Intelligence Research Group and supervised by Dr Aldo Lipani. Check out https://wi.cs.ucl.ac.uk/index.php/people/
Worked on speech and text-based user confidence and expertise detection and measurement in conversational search systems (CSS) using transformer-based multimodal deep neural networks.
Prepared a dataset of user queries named UNSURE from Spotify Podcast, with the use of the word-level transcript to segment out questions.
Crowdsourced using Amazon Mturk service to obtain confidence scores based on the segmented audio files. Performed quality control on collected answers and conducted user agreement analysis.
Trained a text and audio multimodal regression network using pretrained BERT and HuBERT models to predict human confidence scores. Our model showed a human-level performance in confidence score prediction.
Paper published at International Conference on Engineering Applications of Neural Networks (EANN). Check out the paper and the GitHub repo!
Supervisor: Dr Aldo Lipani

University of Cambridge, Department of Engineering, Summer 2021

Successfully developed a web-based interface for Natural Language Processing (NLP) text corpora that enables gender biases to be revealed visually and interactively.
A Flask-based web framework was created where the user can upload their corpora through inputting plain text, URL or txt files. Two NLP algorithms will run, namely the Bias Score Calculation algorithm and the Sentence Parsing algorithm, both based on word embeddings. The user is able to view the Bias scores associated with each token and specific sentence structures. Interactive pivot tables, bar graphs, word clouds, PCA and TSNE graphs are provided for the user to explore and extract information.
The user is also able to input a natural language query, where the query is parsed and the answer is given in the form of a data frame and a bar graph. A debias feature is also available if the user wishes to discard the more extreme parts and retrieve a less biased file.
Supervisor: Dr Marcus Tomalin
GitHub Repo

University of Cambridge, Department of Engineering, Spring 2021

Joined a research group named the Managing Air for Inner Greater Cities (MAGIC) where the broad aim is to develop cities with no air pollution and no heat-island effect. The subgroup directed by Professor Adam Boies targets identifying modes of pollution resulting from different car models.
My job was to identify car plate numbers and hence car models from video footage and hence determine the pollution output from each car model and the pollution that enters the surrounding buildings. I wrote a machine-learning algorithm for detecting UK car plate numbers from videos using computer vision packages such as OpenCV and Pytesseract. Data cleaning was performed using maximum likelihood inference.
Supervisor: Professor Adam Boies
GitHub Repo

SpacePT, Winter 2020

The goal of the project is to identify and predict building damage and forest fires from satellite images. We built an ML-Enhanced Computer Vision Change Detector System for Satellite Images Analysis. We wrote a Convolutional Neural Network (CNN) for differential image detection for satellite building, forest fire and oil spill images. I was the team leader of the unsupervised learning team.
Git Repo not available due to IP rights issues.

University of Hong Kong, Department of Electrical and Electronic Engineering, Summer 2020

Our goal is to generated holography images using deep learning techniques. We used TensorFlow to build various neural networks such as convolutional neural networks (CNN), ResNet, Wide ResNet, DenseNet and SqueezeNet. We performed image to image translations with Generative Adversarial Network (GAN) structures on hologram images.
Supervisor: Professor Edmund Lam
GitHub Repo

University of Singapore, Department of Building, Mar 2017 - Jan 2018

Leader of a student team of three members.
The goal of the project is to produce nano encapsulated organic chemicals for coating. The coating should ideally be self-cleaning (dust-resistant) and thermally insulating. Our Cadmium Orthostannate nanoparticles with 3-(mercaptopropyl)trimethoxysilane suspended in organic solution were shown to have effective thermal insulation abilities, which can save massive amounts of energy if painted on windows.
Silver Award at Singapore Science and Engineering Fair.
Supervisor: Professor Shah Kwok Wei