Bridging Minds and Machines
We study the computational basis of cognition and use insights from cognitive science and neuroscience to build AI systems that are more interpretable, flexible, and grounded in principles of human cognition.
To achieve this, we combine behavioral experiments, real-world developmental data, neural recordings, computational modeling, and machine learning. Specifically, we analyze behavior in memory and learning tasks, study infant egocentric visual experience and eye movements, and examine neural activity from in vitro and in vivo electrophysiological recordings. We then construct neural-level computational models that account for both behavior and neural activity and integrate them into trainable systems that we evaluate on tasks used in animal studies and on machine-learning benchmarks. This work draws on a range of insights from cognitive science and neuroscience, including continual, curriculum-based self-supervised learning from temporal and spatial regularities in the natural world (such as slow-changing sensory inputs), and structured, navigable mental maps of physical and abstract spaces that support memory retrieval and planning.
Recent advances in large language models (LLMs) and vision-language models (VLMs) have created an opportunity to study whether analogous structured representations emerge in foundation models. These comparisons identify where model representations align with human cognition and which inductive biases current AI systems still lack, helping guide the development of more interpretable, sample-efficient, and robust architectures that combine foundation models with cognitive models.
Explore our projects to learn more.