Dileep George @dileeplearning

🎯

Ohh ok I realize that @tyrellturing.bsky.social mentioned evolution. Fine then. But then which neuroscientist believes this?

Hmm…I don’t think it’s impossible.

Evolution could create structures in the brain that are in correspondence with structure in the world.

Clone-structured graph representations enable flexible learning and vicarious evaluation of cognitive maps - Nature Communications Higher-order sequence learning using a structured graph representation - clone-structured cognitive graphs (CSCG) – can explain how the hippocampus learns cognitive maps. CSCG provides novel explanati...

Here's the CSCG paper: www.nature.com/articles/s41...

And here' the CML paper:
www.nature.com/articles/s41...

Good conclusion :-).

somehow chatGPT understand my opinion about successor representations? 4/

I didn't mention partial observability specifically, so it is impressive that this was picked up. Looks like we did something right in our CSCG paper in making this explicit? 3/

It is quite impressive that chatGPT picked up these nuances, picks up a relevant quote from the paper and even emphasizes portions of the response. 2/

This paper turned up on a feed, I was intrigued by it and started reading...

..but then I was quite baffled because our CSCG work seem to have tackled many of these problems in a more general setting and it's not even mentioned!

So I asked ChatGPT... ...I'm impressed by the answer1. 1/🧵

Graph schemas as abstractions for transfer learning, inference, and planning Transferring latent structure from one environment or problem to another is a mechanism by which humans and animals generalize with very little data. Inspired by cognitive and neurobiological insights...

Some of our work could explain this kind of latent graph learning and schema-like abstraction. 2/
arxiv.org/abs/2302.07350

Wow, very cool to see this work from Alla Karpova's lab. She had shown me the results when I visited @hhmijanelia.bsky.social and I was blown away.

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

1/

it takes a bit of getting immersed in the field to know this :-)

This comic would be an inside joke. Many neuroscience papers that study a brain region postulate that the current does a simple transformation of the input to make it easy for brain regions that are 'downstream' to solve the real problem effectively deferring the real problem. ...

𝗛𝗼𝘄 𝘀𝗵𝗼𝘂𝗹𝗱 𝘄𝗲 𝗱𝗲𝗳𝗶𝗻𝗲 𝗮𝗻𝗱 𝗱𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗲 𝗮 𝗯𝗿𝗮𝗶𝗻 𝗿𝗲𝗴𝗶𝗼𝗻'𝘀 𝗶𝗺𝗽𝗼𝗿𝘁𝗮𝗻𝗰𝗲?
We introduce the idea of "importance" in terms of the extent to which a region's signals steer/contribute to brain dynamics as a function of brain state.
Work by @codejoydo.bsky.social
elifesciences.org/reviewed-pre...

It's kinda obvious. #AGIComics has already figured out which brain region is the most important. 😇

And whether top-down't influence is multiplicative or not is very context-dependent. (this is also what is seen in neurobiology). www.science.org/doi/10.1126/...

www.science.org/doi/10.1126/...

Explaining away is a concrete role for feedback computations, and here is one example showing its effect. ....there are many more examples in the paper.
www.science.org/cms/10.1126/...

A detailed theory of thalamic and cortical microcircuits for predictive visual inference A generative vision model offers a detailed theory for cortical columns, thalamus, and computations in cortical circuits.

just want to re-point to this work and remind you that we have proposed and built models that incorporate topdown feedback, and have a theory about it.
www.science.org/doi/10.1126/...

ohh...yes...this is exactly what I think after reading some of the "deep research" reports. ....written by a committee

jumping on the Gemini 2.5 bandwagon... it's an incredible model. really feels like an(other) inflection point. talking to Claude 3.7 feels like talking to a competent colleague who knows about everything, but makes mistakes. Gemini 2.5 feels like talking to a world-class expert with A+ intuitions

already being scaled up...

We could run it to analyze a transformer...

How about running an experiment that runs this process for a known system? Assume the limits of sampling and see what kinds of insights we can get?

I just don’t understand the proposal….what models are going to be the ‘foundation’ for these brain models? Is it the transformer architecture? …because that is the one that is proven to be scalable so far. If so, how are brain insights going to be extracted from a trained transformer?

well, benchmarks are useful, greedy hill climbing on them might lead you to new opportunities even if it doesn’t lead you to new insights. And the job markets can remain irrational longer than you can remain solvent :-).

Give me 10 billion dollars and I’ll do it. 1 billion for developing the hardware and 9 billion to pay for my opportunity cost 😇

Sure. Can you give me a 10billion dollars?

What’s your beef with “processing”? IMO, the specific architectural modifications are about “processing”. The attention circuit is “processing”. No?

Again, tell me which architecture+algorithm you want to scale.