Is the brain mechanistically identifiable?

Nov

2026

3:00 pm – 4:00 pm

In Person (view details)

2026-27 Neuroscience Seminar Series

Description

Dr. Cengiz Pehlevan is an associate professor of Applied Mathematics at Harvard University.

Neuroscience seeks mechanistic explanations of computation. A popular route is to fit dynamical-systems models to tasks or large-scale recordings and read the mechanism off the fitted model. But when can we trust what we read? I will argue that mechanistic identifiability is a central bottleneck. Partial observation and model mismatch allow a data-constrained model to match recorded activity while recovering the wrong circuit mechanism: using temporal integration as an example, I will show how heavy subsampling makes line-attractor and functionally feedforward implementations look deceptively alike to a learner. Rather than betting on a single fitted solution, I will advocate working with families of task-compatible solutions. Iterative Neural Similarity Deflation surfaces alternative solutions by penalizing reuse of previously discovered activity patterns. A second approach introduces a control parameter for how strongly learning restructures recurrent connectivity, yielding a controlled interpolation between reservoir-like and structured regimes and a theory linking disorder and structure to population dynamics, single-neuron statistics, and temporal generalization.

Learn more

Hosted by Dr. Thibaud Taillefumier

To subscribe to our Seminars Series mailing list, please visit https://utlists.utexas.edu/sympa/info/neu-seminars and click on "Subscribe" in the left-hand column.

Please contact Neuro.Seminar.Series@austin.utexas.edu with any questions.