link:

www.biorxiv.org/content/10.1...

and that's a wrap!

We are working on lots of cool related ideas (EEG, DNNs), but for now let me just celebrate and thank my even cooler collaboratores (some of them are here! @hansopdebeeck.bsky.social, @emilyvh.bsky.social, @felipefv.bsky.social).

I am proud of this one!

end/n

This multi-lev approach (WHAT, HOW, WHERE) can be directly applied to existing neuro and DNNs datasets, and suggests a general principle:

expertise emerges when the brain discovers the latent structure of a domain, and reshapes its representational space to express that structure efficiently.

11/n

Overall, we find that expertise reorganizes the brain at three levels:

1️⃣ WHAT is represented → relational structure > appearance
2️⃣ HOW it is represented → compact, low-dimensional manifolds aligned with task structure
3️⃣ WHERE it is represented → from domain-specific to domain-general networks

10/n

This reveals what we call NETWORK–CODE INVERSION:

- Novices use domain-specific networks to produce relatively unstructured, domain-general (not tuned to the task) codes.

- Experts repurpose domain-general control networks to encode highly structured, domain-specific knowledge.

9/n

3️⃣ WHERE do these optim codes live?

🧠 Experts rely more on
- working-memory systems
- navigation-related regions
- memory-retrieval networks

👁️ Novices rely more on
- early visual cortex
- face/object regions
- language areas

👉 A shift from domain-specific → domain-general control networks.

8/n

2️⃣ HOW are representations structured?

Using manifold dimensionality (Participation Ratio), we find lower-dimensional, more compressed neural codes in experts. And these compressed manifolds carry more task-relevant information.

👉 Experts pack more information into fewer dimensions.

7/n

If we want to understand how learning reshapes representations, we must look at their structure.

This brings us to our second question..

6/n

But representing the right information is only part of the story.

Previous psych theories (chunking, templates) and computational frameworks (multiplexed representations, manifold learning, rich vs. lazy coding) all suggest a deep link between learning and representational compression.

5/n

WHAT is represented? -- brain

Brain–model RSA shows the same shift: both groups encode visual features, but only experts encode high-level, task-relevant structure.

In other words,

👉 expertise changes WHAT is represented: from low-level, surface features to high-level, relational structure.

4/n

1️⃣ WHAT is represented? -- behavior

Behavioral RSA shows that experts organize their value judgments around relational and goal-relevant structure. Visual similarity barely plays a role.

Novices, meanwhile, show much less structured preferences.

3/n

To test this, we built a stimulus set spanning low-level (visual similarity) to high-level (strategy, checkmate) structure, and turned these into RDMs capturing our theoretical models.

**Enter: Representational Similarity Analysis (RSA).**

2/n

I see the point, but that might be a too simplistic example. In brains it is unlikely that a single neuron "does the job", and from a representational perspective what we call "noise" may still be a behaviorally relevant signal at the pop level (stim X shows property i but not j).

Congrats, Martin! Compelling results and very interesting methods. And, even more exciting, your conclusions/results are in line with our (now published, link in my last post) work on foveal feedback.

Looking forward to discussing this further!

And that’s a wrap! 🎉 Huge thanks to my co-authors, reviewers, and everyone who helped with this project!

The full paper is open access: www.sciencedirect.com/science/arti...

And feel free to reach out if you have any questions😁