Research focus

01   Selective attention and adaptive computation

We investigate architectures that allocate computation according to the information and task at hand. Sparse attention, latent representations, and dynamic routing offer ways to study which signals a model selects, how it combines them, and when additional computation is useful.

02   Learning under changing conditions

We ask how a model can incorporate new information while retaining useful prior knowledge. Our interests include continual learning, distribution shifts, and interactions among learning objectives, with an emphasis on the mechanisms that support or disrupt adaptation.

03   Compositional and reusable knowledge

We explore representations that separate and recombine useful components of a task or experience. By connecting modularity with flexible information integration, we investigate how learned knowledge might support unfamiliar combinations, contexts, and goals.

How we investigate

We translate a candidate computational principle into a model, test the mechanism through controlled comparisons, and examine how its behavior changes across tasks and conditions. Comparisons with neural or behavioral observations help refine the underlying hypothesis. We assess adaptation, computational efficiency, and interpretability without assuming that biological inspiration alone establishes a match to the brain.

Across our research framework

This direction connects neuroscience questions with new AI architectures. The resulting ideas can shape brain foundation models, while model behavior generates hypotheses for neural representation research and adaptive closed-loop systems.

Continual learning · Sparse attention · Compositionality · Adaptive computation