Projects Highlight
Multi-sourced Event Detection and Multi-Dimensional Analysis based on Event Cube
Targeting suicide detection of Hong Kong youngsters, this project aims to provide insight on real-world event detection and prediction underlying the Big Data.
Big Data-Driven Airport Resource Management (BigARM)
Develop big data-driven approaches to solve resource optimization and management problems in airports.
- Flight arrival time prediction: A cost-weighted matrix factorization algorithm to fuse cross-domain data
- Arrival bag count prediction: A weighted moving linear regression algorithm to discover sequential patterns
- Load-balanced resource allocation: An incremental reinforcement learning algorithm and a recursive balanced k-subset sum partition algorithm
A Distributed Approach Pioneering the Edge GenAI Revolution
A decentralised paradigm that targets edge-based GenAI for specialised reasoning and planning applications:
- Single-step Distillation - an innovative approach that transforms complex pretrained models into efficient generators
- Novel “Local-feature Enrichment and Global-content Orchestration” (LEGO) - Noteworthy 60% reduction in sampling time compared to state-of-the-arts
- InfiAgent framework - comparable to OpenAI‘s Code Interpreter
- InfiMM - a top-ranked, smaller multimodal model (7B, 13B)
What Strike the String of Your Heart? --- Music Therapy for Anxiety by Machine-composed Music
- Music can evoke strong emotion and music appreciation is helpful to release anxiety.
- This project aims to enable computer generate therapeutic music like human composers.
- Artificial intelligence techniques including deep learning, EEG-based brain imaging, and algorithmic composition are developed.
High-quality Long Video Generation Towards Multi-Scene Scenario
- The first research team to propose the new concept of harmonious generation of multiple motion patterns.
- The main innovation is the design of an effective algorithm for adaptive learning of prior knowledge of motion, used to constrain the spatial domain attention and temporal continuity of multiple motion patterns.
- Using the physical principles based on object motion and the continuity of object motion, a propagation model for the motion parameters of objects in adjacent frames and long videos for effective video generation has been designed.
Next Generation Digital Camera Imaging: Human Centered Image Reconstruction, Enhancement and Evaluation
This project aims to develop next generation human-centered digital camera imaging technologies, including a unified end-to-end pipeline for in-camera image reconstruction, a set of sophisticated deep image enhancement techniques, and human-centered indices for image quality/aesthetic evaluation. The human-centered data driven feature of our technology can form a positive loop between the users’ feedback and product development, which can constantly improve the user experience very efficiently.
Practical Post-Quantum Zero-Knowledge Proofs
- A zero-knowledge proof (ZKP) allows one party to convince another party the validity of a statement without revealing any additional information
- Useful for (1) Privacy Protection; and (2) Speed up repeated verification
- This project aims to develop ZKP that is (1) practical (i.e., efficient and deployable on existing hardware) and (2) highly secure (i.e., security even against powerful quantum attackers)
Securing Smart Contracts and the Runtime of Blockchains
The support of smart contracts empowers running various applications on the blockchain other than cryptocurrency. Unfortunately, vulnerable smart contracts and their execution environments can be exploited to cause financial damage to users. We aim at designing new approaches to expose the security and performance issues in smart contracts and their runtime and to improve them.
Enhancing Digital Asset Security Based on the Blockchain Technology
The blockchain technology has the property of decentralization, data transparency and immutability. We conduct innovative research tasks in four aspects. (1) Secure cloud data search service. (2) Anti-attack digital asset trading system. (3) Digital asset exchange traceability and privacy protection. (4) Reliable and secure decentralized digital asset storage. The project will generate long-term impact on digital asset security, and its results can be directly applied to industrial products.
BaiTest: A Platform for AI Evaluation in Smart Buildings
Machine learning (ML) models have been widely developed for building HVAC systems. Intrinsically, there has been a lack of a methodology related to building ML model evaluation. In this project, we proposed a BaiTest (Building AI Test), a new evaluation methodology for ML modelling in buildings with an evaluation platform to realise our methodology. With AI forecasting models provided by the Electrical & Mechanical Services Department of the HKSAR Government, a BaiTest can be used by building operators and AI developers to compare and select appropriate ML models through interactive visualisation services.
NeRF2: Neural Radio-Frequency Radiance Fields
Over 150 years after James Clerk Maxwell's discoveries in electromagnetism, challenges still exist in predicting how radio signals travel through complex environments. Our research introduces the Neural Radio-Frequency Radiance Field (NeRF2), a breakthrough model that simplifies understanding of radio signal paths, even in cluttered settings. This tool could eventually enhance technologies like indoor localisation and 5G networks, making connections faster and more reliable. Our work won "Best Paper Award - Runner-up" at ACM MobiCom 2023.
A System Identification based Oracle for Control-CPS Software Fault Localization
We propose an oracle based on the well adopted autoregressive system identification (AR-SI). With proven success for controlling black-box physical systems, AR-SI is adapted by us to identify the buggy control-CPS as a black-box.
