Research
Our research aims to uncover the fundamental principles underlying learning and computation in the brain. We work at the intersection of neuroscience and artificial intelligence (AI). While modern AI demonstrates superhuman capabilities in domains ranging from mathematics to the arts, it remains hindered by massive energy demands, a reliance on enormous datasets, and an inability to learn continuously. In contrast, the human brain is capable of flexible reasoning from sparse experience and adaptation across a lifetime while operating on approximately 20 watts. We develop theoretical frameworks to understand how synaptic and neural plasticity, intricate connectivity patterns, and closed-loop environmental interactions enable the brain’s remarkable efficiency. Our ultimate goal is to translate these biological insights into the next generation of intelligent algorithms.
Ongoing Projects
Online Continual Learning in Brains and Machines
The brain acquires new skills and knowledge seamlessly from a continuous stream of information. This sample-by-sample learning remains a major challenge for artificial neural networks, which typically require static, repetitive training to avoid forgetting. By integrating deep learning theory, computational modeling, neural data analysis, and behavioral psychology, we investigate the fundamental principles of online learning. Our goal is to uncover the mechanisms of lifelong adaptation in biological systems and develop algorithms that accelerate training and efficiency in both machines and brains.
Normative Theory of Neural Connectivity and Architectures
Recent advances in experimental techniques have provided unprecedented access to micro- and macroscopic neural connectivity data. While these data are often viewed as constraints for models, they may reveal deeper functional principles. We develop normative theories, mathematical frameworks grounded in statistical learning theory, Bayesian statistics, and statistical physics, to explain why the brain is wired the way it is. These frameworks allow us to predict how specific connectivity patterns support learning and complex neural computations.
High-Dimensional Neural and Behavioral Data Analysis
Modern neuroscience generates vast, high-dimensional datasets that challenge traditional analytical methods. We examine how high dimensionality introduces bias into conventional data metrics, and we develop domain-specific approaches to extract meaning from complex data. Our current methodological work focuses on extracting latent structures underlying individual variability in animal behavior.
Structured Knowledge Processing
The mammalian brain excels at organizing information into structured knowledge, such as spatial maps and relational hierarchies. However, we still do not fully understand how the brain constructs these representations in memory or leverages them for rapid inference. By combining modeling with experimental data, we study how biological systems manipulate structured knowledge. This work informs the design of new AI architectures with more flexible, human-like reasoning capabilities.