Raia Hadsell is a vice president of research at Google DeepMind and co-leader of its Frontier AI unit. Her work has helped shape contrastive representation learning, robots that perceive their surroundings, neural networks that retain previously acquired knowledge, and interactive worlds for training intelligent systems.
Hadsell studied religion and philosophy at Reed College before pursuing computer science at New York University, where she completed doctoral research with Yann LeCun on Siamese neural networks and robotic perception. Her dissertation explored long-range vision for off-road robots and received an outstanding dissertation award in 2009. That early work on contrastive loss anticipated techniques now central to embeddings, image recognition, and multimodal retrieval.
She subsequently worked at Carnegie Mellon University’s Robotics Institute and SRI International’s Vision and Robotics group before joining DeepMind in 2014. There, her research expanded into reinforcement learning, navigation, transfer learning, and systems that acquire multiple capabilities over time. Her contributions include policy distillation, progressive neural networks, and elastic weight consolidation.
Ideas shaping her research
- Learning without catastrophic forgetting. Hadsell co-authored research on preserving previously learned tasks by selectively protecting parameters important to existing capabilities while allowing others to adapt. This approach to continual learning addresses a basic requirement for agents operating across successive tasks: gaining new abilities without erasing old ones.
- Retrieval needs representations that cross modalities. Her early work on Siamese networks extends into embedding systems that place text, images, audio, and video within a shared semantic space. She treats retrieval as an essential complement to generation and highlights Matryoshka Representation Learning, which permits efficient lower-dimensional searches while retaining the option of richer representations.
- Forecasting must capture uncertainty. Hadsell has advanced Google DeepMind’s work on weather prediction, including GraphCast’s graph-based atmospheric modeling and GenCast’s probabilistic forecasts. Modeling the range of plausible outcomes matters particularly for storms and other hazardous conditions, where uncertainty directly affects preparation.
- Interactive worlds can train embodied intelligence. The Genie world models build on her background in robotics and reinforcement learning, progressing from short two-dimensional game environments toward persistent three-dimensional worlds responsive to user actions and new prompts. Such environments could support gaming and education while giving agents opportunities to practice perception, planning, and movement. Her frontier AI research overview connects these systems with a broader ambition: intelligence that understands physical environments as well as language.
In 2022, Hadsell co-founded Transactions on Machine Learning Research with Kyunghyun Cho and Hugo Larochelle, establishing an open journal with continuous publication and transparent review. She served as a founding editor-in-chief through 2023 and is also a UK government AI ambassador.
