What we work on

We design decentralized and federated learning protocols that work without a central coordinator, studying their robustness to node and data disruption and how network topology shapes their performance.

We bring model training and inference onto resource-constrained and mobile devices, addressing the systems and algorithmic challenges of learning close to where data is produced.

We build mathematical and LLM-based models of human behavior — social search and expert finding, personality prediction from text — to understand how online and offline social systems actually work, not just to build AI on top of them.

We study the structure and dynamics of complex networks — signed relationships, ego networks, and opinion dynamics — to understand how influence and information propagate through social and communication graphs.

We work on causal structure learning under real-world constraints, including budgeted-intervention methods for discovering cause-effect relationships in pervasive and social systems.

We study how social interaction and network structure change in immersive virtual environments, including AI-driven independent avatars that can manage social interaction on a user's behalf.

We study how humans and AI agents can work together effectively, including neurosymbolic approaches that combine neural networks with symbolic reasoning in human-in-the-loop settings.

We work on communication and coordination between devices that connect intermittently and without fixed infrastructure, a setting we also refer to as the Internet of People.