Learning Across Systems and Scales

Shared structure, different systems

A changing system can be observed at many resolutions, from individual measurements to the behavior of a whole network. Each view reveals some relationships and conceals others.

We investigate whether temporal representations can connect these views while preserving the context that gives each observation its meaning.

Transfer with boundaries

Patterns learned in one domain may not survive in another. Useful transfer needs to distinguish reusable dynamics from relationships that depend on local conditions and measurement choices.

Testing the limits of transfer

We would evaluate on held-out domains, sampling frequencies and system sizes, comparing transfer with models trained in the target setting. These comparisons should reveal both reusable dynamics and the conditions that must be learned again.