Our lab’s research spans the full stack of machine learning systems, but a central focus today is making large-scale pre-training of foundation models radically more efficient. Modern LLMs demand enormous communication bandwidth and tightly integrated datacenters, limiting who can train them and constraining the pace of innovation. We develop new optimization methods, communication-efficient training strategies, and architectural techniques that reduce these requirements, enabling large models to be trained across heterogeneous, distributed, and bandwidth-constrained hardware—including GPUs separated by geographic distance. Our goal is to rethink pre-training so that high-quality foundation models can be built sustainably, affordably, and in a wider range of computational environments.

Alongside this flagship work, the lab continues to advance a broad portfolio of research at the intersection of machine learning and systems. We develop privacy-preserving and federated learning frameworks capable of operating on sensitive or distributed datasets; design efficient on-device and embedded ML systems that bring intelligence closer to users and edge hardware; and contribute to emerging issues in AI trustworthiness, compliance, and unlearning, where reliable system behaviour and regulatory alignment are essential. Across these domains, we emphasize principled algorithmic development together with practical systems design, enabling ML models to train, adapt, and operate effectively in real-world conditions.

Collectively, our research aims to build the next generation of machine learning systems, from globally distributed pre-training pipelines to reliable, privacy-preserving edge intelligence, that are both high-performing and fundamentally more efficient, trustworthy, and accessible.

A list of publications can be found on our publications page. It may also be useful to check group members’ Google Scholar pages.

For certain research projects we have produced blog posts that expand upon the intuitions and ideas in their corresponding publications. A list of posts (ordered chronologically) can be found below: