Albert Albert Lopez Serrano

Compositional Generalization, that is, the ability to systematically recombine previously acquired knowledge to solve new unseen tasks in structured environments (robotics, language, image understanding, ...), has been a long-pursued capability of intelligent autonomous systems. The gap between human intelligence's exhibited systematic composition and recombination of acquired knowledge and the performance of learnable models has been a long standing challenge in the AI community, with roots spanning to the origin of the field. While language-based foundational models have recently showed some success on this front, the problem is still lacking a principled theory dealing with the phenomenon, the necessary conditions for it to arise, and specific tools to guarantee that capability emergence beyond extensive training schemes on vast data collections. This problem, and the ability to predict and prove its failure modes is critical in environments where failure is particularly costly, i.e. robotics and autonomous systems. The rise of foundational models for robotics, often intertwined with monolithic large language engines, has highlighted the need for methods to precisely assess the reliability of compositional generalization capabilities in AI systems, as well as techniques to assess when, how and why these systems might fail to generalize behaviours in novel environments or tasks involving simultaneous and sequential systematic combination of pre-acquired behaviours.

In this project we will aim to devise a principled theory of compositional generalization in AI systems, with a focus on actionable techniques to elicit such systematic composition, provable guarantees of their performance and assessment of their shortcomings. Leveraging the techniques and expertise of the advisory team on Bayesian Deep Learning, Information Geometry and Reinforcement Learning, we will target foundational questions in the topic with an emphasis on reliability, computable theoretical guarantees and connections to state-of-the-art solutions for autonomous AI systems.

Distinctively, we aim to move from train time to test time adaptations based on underlying task structure, expecting to get solid certificates on top of the existing models, thus providing the community with more amenable, easier to probe certificates. This focus on test-time adaptation, will allow to overcome costly pre-training and extensive data-only approaches. We expect to devise methods and theory for uncertainty-aware inference-time reasoning for physical intelligence leveraging intrinsic symmetries in the data-generating processes.