What we work on

Machine learning and signal processing on graphs

Much of today's data lives on networks: sensor networks, molecules, transport systems, social and biological interactions. We design graph neural networks and graph signal processing tools that exploit this structure while remaining principled and interpretable.

Current directions include learning over temporal and continuous-time dynamic graphs, causal learning via directed acyclic graphs, learning from multi-domain data with product graphs, and conformal inference that attaches reliable uncertainty to graph predictions.

Topological signal processing and learning

Pairwise graphs cannot capture interactions among groups of entities. We work with higher-order domains such as simplicial complexes, cell complexes and hypergraphs, and study signals that live on edges, triangles and beyond.

Our work covers sampling and recovery of signals on simplicial and cell complexes, self-supervised simplicial representation learning, Gaussian processes for edge-flow prediction and simplicial closure, and topological scattering networks.

Signal processing and AI for communications and sensing

Next-generation wireless systems will sense their environment as well as communicate. We develop signal processing and learning methods for integrated sensing and communications (ISAC), reconfigurable intelligent surfaces (RIS), near-field communications, and distributed sensing and communication networks.

Examples include beamforming and precoder design for RIS-assisted and distributed ISAC, learning-based and sparse precoders, sparse array design, channel estimation with one-bit spatial sigma-delta converters, and radar imaging with RIS.

Decentralized and federated learning

When data is spread across devices that cannot share it, learning has to happen collaboratively. We design decentralized and federated optimization algorithms that handle constraints, non-IID data, privacy requirements, and limited communication.

Our work centres on projection-free (Frank–Wolfe and conditional-gradient) methods, differentially private decentralized learning, communication-efficient schemes with learned quantizers, and algorithm unrolling.

Sparse sensing and sampling

Sensing is expensive in hardware, energy and bandwidth. Sparse sensing asks which measurements to collect so that an inference task still succeeds. This theme runs through the lab's work, from sensor placement to graph sampling.

Topics include sensor selection for estimation and detection, submodular and convex sensor placement, sparse arrays, sampling of graph signals and covariance, array calibration, and localization. The monograph Sparse Sensing for Statistical Inference (Foundations and Trends in Signal Processing, 2016) gives an overview.