CCN logo CCN 2026 · K&Ts
New York University · August 3–6, 2026

Adaptive, degenerate,
and yet comparable?

Rethinking representational comparisons

Erin GrantUniversity of Alberta Lukas BraunAllen Institute Eleanor HoltonPrinceton → Columbia Marvin TheissUniversity of Tübingen
Erin Grant Lukas Braun Eleanor Holton Marvin Theiss

Monday, August 3, 4:30–6:15 pm · New York University

Neural networks adapt their representations to task demands, yet many distinct configurations can implement the same behavior. This keynote and tutorial (K&T) establishes when representational comparisons permit valid inferences about computation by accounting for the adaptive and degenerate nature of neural representations. We make the case for representational pluralism, that multiple neural representational geometries can support the same task performance while differing in their downstream computational affordances.

View on the CCN 2026 program
keynote

Adaptive, degenerate, and yet comparable?

Neural network theory delineates when representational geometry can and cannot support inferences about computation.

 tutorial 1

Linear networks

Function and representation are dissociable in linear networks.

Hands-on tutorial tutorial 2

Nonlinear networks

Function and representation are dissociable in nonlinear networks.

Interactive slides tutorial 3

Representational pluralism

Networks solve identical tasks with different representational geometries.

Hands-on tutorial