Our research basically spans across three different, but integrated, topics:  

1.
graph signal processing (GSP) and graph learning (for non-Euclidean data)
- graph based signal power prediction
- 2nd-order online graph learning with application to learning causal graph for atrial fibrillation (for prevention of stroke, arrhythmia, death)
- fast and robust 2nd-order online graph learning (improvement on the above)
- we have just started to explore graph (convolutional) neural networks (GNN/GCNN) with the goal doing data analysis and communication problems

2.
distributed and federated learning (DL and FL)
(This work was actually inspired by other automotive radar partner WNC Corp in Taiwan who wanted to design a classification system using their 3D radar)
- we are continuing our work in supervised and unsupervised distributed and federated learning for heterogeneous networks (i.e. with statistical and system (asynchronicity) heterogeneity)
 - designing distributed optimization algorithm to handle learning in heterogeneous networks
- with applications to channel estimation and beamformer design for intelligent reflective (IRS) assisted system (see below) 
- I am interested in using FL and/or DL for training Large Graph (Foundation) Models for different tasks
- I am interested in looking at the effects of statistical and system heterogeneity during FL using generative models such as GANs, VAEs, and diffusion probability models
- Online federated and distributed learning (data streaming)
- I am particularly interested in looking into multitask learning in the distributed and federated framework and how statistical and system heterogeneity affects the learning process

3.
6G communications using deep learning
a) practical intelligent reflective surface (IRS)
- focusing on channel estimation, joint (robust) beamformer design and aerial IRS (AIRS) placement by considering practical issues at the antenna array and objective functions
 - solutions involve model-driven neural network, graph convolution neural networks, policy gradient methods (from deep reinforcement learning)
b) robust beamfocusing for holographic (large-scale) MIMO antenna (HMA) array with mutual coupling and near-field effects.
   - using diffusion model for channel modeling
   - joint beamforming/beamfocusing and diffusion model channel model design
c) low latency and low energy consumption virtual function network design using service chain graphs (related to above work on GNN and GCNN) using reinforcement learning and diffusion optimization solver
d) federated generative AI learning for end-to-end communications
e) network orchestration problem and using genAI model to solve different (cloud-edge) communications and networking problem