Research
Approach and Interests
The Signal
Processing and AI Group (SIPAI Group) designs numerical optimization algorithms
to solve different machine/deep learning problems and apply the solutions to different
applications, such as data association for automotive radar, channel estimation
and precoder design for wireless communications, data analysis using
distributed and federated learning, and graph signal processing and graph
learning for communications and biomedical engineering.
·
GSP
and graph learning
o We
have been focusing on online
graph learning which tracks on time-varying graph that models non-Euclidean
data that lie on an irregular structure. Applications include modeling of
cyberattacks, brain, heart, and network traffic, to predict, e.g. the onset of
atrial fibrillation (AF).
o We
are also investigating the problem of (online) attribute graph clustering using
network rolling, graph neural networks, and model agnostic meta learning
(MAML), with application to community detection, recommendation systems and
fraud detection.
·
Distributed
and federated learning (DL and FL) in Heterogeneous Networks
o We
have been looking at the problem of supervised, self-supervised, and unsupervised
distributed and federated learning, with the first two aiming to train MLPs and
model-based deep learning models in heterogeneous networks (with statistical
and system heterogeneity)
o We
are also looking how optimizing communications components can assist in
minimizing the impact of asynchronicity in FL
o Distributed
online graph learning problem becomes essential to prevent raw data from being
transmitted to the cloud for learning task
·
Signal
processing for 6G communications
o This
work has mainly been focused on developing physics-informed neural networks
(PINNs) for solving XL-MIMO/HMIMO channel estimation and precoder/beamformer
design for 6G systems and intelligent reflective surface-assisted
(IRS-assisted) systems by considering nonidealities in the antenna array near-field
effects, and polarization. Focus will be
on the FR2 (24 – 70 GHz) and FR3 7 – 24 GHz bands
o We
also looking at the feasibility of using LLMs and diffusion models to solve
optimization models in communications