Research

Thrust 1 Core AI Research

A few years ago, the first Convolutional Neural Network (CNN) surpassed human performance on ImageNet. However, despite achieving superhuman accuracy on benchmarks with ideal data, existing machine learning models struggle to provide reliable predictions in unseen or adverse viewing conditions, such as low light or bad weather.

The achievements of deep learning heavily rely on extensive training data. Unfortunately, real-world data often diverge from the distribution observed in the training set, leading to a notable decline in performance. Consequently, a significant barrier to the widespread deployment of machines "in the wild" lies in the adaptation of off-the-shelf models to handle adverse conditions when faced with limited data. 

My research revolves around unveiling the inner workings of neural networks and delving into the latent representations within them. This pursuit is crucial in addressing the aforementioned challenge. I have developed various theoretical techniques to visualise and interpret these latent features, subsequently aligning them with prior knowledge derived from the physical world. In this endeavour, I have proposed several manifold-based and self-supervised learning algorithms, among others, which have emerged as state-of-the-art benchmarks within the research community.  

Thrust 2 Catalyse Diverse Research through AI Collaboration 

I am particularly interested in collaborating with specialists from various domains to leverage artificial intelligence for their research endeavours. My approach goes beyond using AI as a mere tool; I'm committed to constructing intricate computer models to learn and align latent representations with domain-specific prior knowledge. 

My research spectrum encompasses an array of applications, such as providing diagnostic support within the medical field, advancing terahertz imaging techniques, contributing to the realm of nuclear fusion, delving into the intricacies of neuron science, exploring mental disorder diagnostics, and more. Through these multifaceted pursuits, I aspire to bring the transformative potential of AI to the forefront of diverse disciplines. 

Among these fields, the medical domain is significant in my research endeavours. My research covers various modalities and medical conditions, including breast cancer, brain tumours, chest X-rays, cardiology, etc. Diverging from the conventional approach of introducing minor algorithmic tweaks and technical modifications, my primary aim is to drive conceptual breakthroughs in methodologies. This overarching goal centres on elevating the practical diagnostic process to a level that significantly improves patient well-being. 

A prime example of my approach is the collaborative effort with institutions like NIH and the National Cancer Center. Together, we established a groundbreaking milestone by constructing the world's first baseline for multi-modal Large Language Models (LLMs) in clinical diagnosis. 

Furthermore, I collaborate with specialists from various disciplines, such as anthropology, material design, political science, and beyond.