How can we make progress in developing a general model of neural computation rather than a series of disjointed models tied to specific experimental circumstances, ask Eva Dyer and @tyrellturing.bsky.social in the latest entry in our NeuroAI series. www.thetransmitter.org/neuroai/acce... - ThreadSky

thetransmitter.bsky.social • 38 days ago

How can we make progress in developing a general model of neural computation rather than a series of disjointed models tied to specific experimental circumstances, ask Eva Dyer and @tyrellturing.bsky.social in the latest entry in our NeuroAI series.

https://www.thetransmitter.org/neuroai/accepting-the-bitter-lesson-and-embracing-the-brains-complexity/?utm_source=bluesky&utm_medium=org-social&utm_campaign=20250326-embracing-brains-complexity

Comments

danieljbutler.bsky.social•37 days ago

Much more useful than a massive predictive model of raw brain activity would be an LLM trained to ask good questions about brain data, and then try to answer them.

neuralreckoning.bsky.social•38 days ago

My thoughts on this:

https://bsky.app/profile/neuralreckoning.bsky.social/post/3llcgyhv5xs2u

pessoabrain.bsky.social•37 days ago

I just don't think we can sample the neural-behavioral space to really push things forward in terms of "understanding" as everyone is pointing out. But I'm trying to be more positive than usual so a CERN like project would be fascinating hypothetically.

jbarbosa.org•37 days ago

If we can't sample that space, how are we training models to predict it?

pessoabrain.bsky.social•37 days ago

Sufficiently sample, I should say.
Imagine that we want to understand the entire set of oceans and their entire behavior, dynamics, etc with all species in there.
Then investigating an aquarium sized object won't get you very far.

jbarbosa.org•37 days ago

With a foundation model, no, it would not work at all outside the small aquarium. With physics from first principles we would do much better.

neuralreckoning.bsky.social•37 days ago

Yes exactly this! Exceptionally well put @jbarbosa.org 👏

rdgao.bsky.social•37 days ago

this here is the hubris I was talking about.

tyrellturing.bsky.social•37 days ago

Who says we can't sample the space we want?

jbarbosa.org•37 days ago

Luiz said it. And I agree, at least superficially. Seems too high dimensional

tyrellturing.bsky.social•37 days ago

I'm not convinced. The space of human language and images are pretty high-dimensional, yet we have models that can develop good latent representations of them.

gmckcypress.bsky.social•38 days ago

David Sussillo’s response captures the key point. AI can make accurate predictions, but those are infamously impossible to explain. “Understanding” is nowhere to be found in AI results.

Understanding brains will come from following Dobzhansky: Nothing in biology makes sense except via evolution.

nicholdav.bsky.social•36 days ago

💯

https://bsky.app/profile/jmxpearson.bsky.social/post/3lc4zwyckuc2x

andrewtanyongyi.bsky.social•37 days ago

"What are the barriers that prevent us from pursuing an AlphaFold for the brain?:

In AlphaFold the underlying true physical premise is that sequence predicts structure. What is the analogous premise for the analogous systems neuroscience project?

Comments

Posting Rules

Reply