Research
Our research spans fundamental AI/ML methods and their applications in real-world domains. We aim to build intelligent systems that are data-efficient, trustworthy, and impactful.
Few-Shot & Zero-Shot Learning
Developing methods that enable machine learning models to generalize from very few or even zero labeled examples. Our work explores meta-learning, transfer learning, and prompt-based techniques to achieve strong performance in data-scarce settings.
Deep Reinforcement Learning
Building intelligent agents that learn to make sequential decisions through trial and error. We investigate model-based RL, multi-agent systems, and safe RL with applications in robotics, autonomous navigation, and resource optimization.
Self-Supervised & Generative Learning
Exploring contrastive learning, masked autoencoders, and diffusion models for learning rich representations without manual labels. Applications include image generation, video understanding, and scientific data analysis.
Brain-Computer Interface
Decoding neural signals (EEG, fMRI) using deep learning for understanding brain activity and enabling direct brain-to-device communication. Our work bridges neuroscience and AI for assistive technology and cognitive computing.
Recommender Systems
Advancing personalized recommendation through causal inference, knowledge graphs, and multi-modal understanding. We focus on fairness, explainability, and cross-domain generalization in recommendation scenarios.
IoT & Intelligent Transportation
Designing AI solutions for smart city applications including traffic prediction, sensor data fusion, and edge computing. We develop efficient models that operate under real-world resource constraints.
Causal AI
As co-theme leader of the Responsible AI Research Centre (Theme 4: Causal AI), we develop methods for causal discovery, causal reasoning, and interventional modeling to build more trustworthy and interpretable AI systems.
Computer Vision
Applying deep learning to visual recognition tasks including object detection, scene understanding, and action recognition. Our work emphasizes sample-efficient learning and domain adaptation for real-world deployment.