Proving and improving the Platonic Representation Hypothesis with JEPAs

Seems to me that it matters because it has implications for how we can train smaller models?

It’s less profound sounding if instead of talking about “representations of the world” you talk about compressed representations of data or decision boundaries in a classifier over data.

What if we ultimately find there are two different, internally consistent but mutually incompatible, representations?

What if one of them is a Republican and one is a Democrat? We already see this in some systems today.

More worrying, what if one is (for lack of a better word, what we might describe as) God and one is the Devil?

The Platonic Representation Hypothesis initially feels profound but actually seems inevitable after some thought because all neural networks are trying to model the same reality. It would actually be weirder if it wasn't true and neural networks could form completely incompatible representations of the same reality. In some sense it seems similar to reducing different math equations to the same most simplified form

Haven’t thought too deeply about this, but as a physicist we’d do things like switch from Euclidean to Spherical coordinates or a completely different dynamical representation like the Langrangian, and this makes certain problems dissolve (or at least become much simpler to solve analytically).

Does it even make sense to claim there is a “platonic” / true representation? Maybe there is an optimal one for a specific task, eg representing a small map is great in the Cartesian frame. But you need a spheroid instead of a plane to accurately model the earth, so the latter is “more real” in some sense. But these seem like engineering tradeoffs of error vs compute and a given implementation could correctly pick either depending on its constraints.

Edit: reading the paper more I think this is a “RTFA” point, their “PRH” is a very explicitly defined mathematical theorem and as such my hand-wavey thoughts above don’t really apply.

Yeah, it's neither spooky not really that profound. Perhaps a more layman-friendly explanation:

A good encoder has to capture real relationships between things in the world (what’s similar to what, what goes with what). Since there’s only one world, different good encoders converge to the same relational structure, which shows up as geometric similarity between their embedding spaces.

> Success of autoregressive next token prediction feels like a stroke of luck.

Hard disagree. Self-attention + MLP was a godsend because it keeps scaling with compute ("the curve goes up and up and up"). Autoregressive next-token prediction turned out to be the most compute efficient way to train it.

We are almost 10 years into this, and it's still here (and in a field that moves at light speed!).

This is not a stroke of luck. It's here to stay.

> Im bullish on jepa models.

Eh, dunno. They've shown lots of promise "in the lab", on these narrow domains shown in papers, but AFAIU they're hard to train on some domains (i.e. language). Someone would have released something if it's that promising. Even RWKV had more models trained that showed real comparisons. Jepa seems like a theoretical toy, until they actually prove that it works in real world (hah!) models...

Consider a binary tree of concepts from the root level more abstract concept to derived concepts applicable to specific domains. Do there exists additional derived concepts which are like an oblique plane which intersects the tree. Sort of like a concept which traveses maths, physics, biology, chemistry and intersects at different abstraction points. I am definitely sure language is flexible enough to have such entities.

Is theft a reducable concept?

I've been wondering on this awhile if generalized models can exist if certain prime concepts don't exist in them? Quite often people say we should leave things like the robot apocalypse out of training data so AI doesn't have the idea of rising up, but personally I don't think that will work in a generalized system. It's not a prime concept itself. You'd have to remove so many prime concepts that lead to that the model would be so narrow it wouldn't be useful for things we want them to be.

The technical term you’re reaching for is an ontological primitive

That's also almost certainly not a cat.

Their method uses two models trained separately on images and captions and turns images into captions without ever having seen a captioned image. So of course it's not very good at the task. Almost all the results in the paper are bad in an absolute sense. But they're surprisingly good considering the enormous handicap of not having any paired training data. So this might be not completely useless in settings where paired data is actually hard to come by.

The trick is to understand what is going on here: rather than a singular model trained for this exact job, this is performed by a awful hackjob of two models cut open and glued together at the seam. That glue layer is a single matrix, a fully connected layer, that is the only thing re-trained in the combined model. The rest of the coefficients are fixed. Naively, this shouldn't work - the internal representations of each model are unconstrained and should intuitively have completely different shapes. But this paper shows that this isn't quite the case: the difference between the latent spaces can be mostly expressed in a single matrix transform. And that allows that merged model to work at all.

It's clickbait. They don't do anything like prove the platonic representation hypothesis.

They do seem to have some really interesting training efficiency improvements, though. The paper is at least worth a read.

Concepts emerge as good compressions of data, or repeated patterns, or recurring distinctions that prove useful to name and recognize.

Text and image often represent same relationships: e.g. you can see squirrel near a tree in an image, and "squirrel" and "tree" are often mentioned together. So both image and text encoders would see concepts of "tree" and "squirrel" close in some sense.

The easiest way to make JEPA and LLM to work together is to map JEPA encoding into input embedding, or inject it to a residual stream - perhaps via cross-attention.

But I don't think you can train JEPA and LLM together: JEPA requires a different training regime. Way easier to take already-trained JEPA and integrate it into LLM. The later can be like 30 minute experiment AI agent can code for you.

The results are in the math. Does not really matter the terms used.

You wont get anything materially useful, apart from an intuition, from using different terminology.

I share your exasperation, it is especially tiring to see that what would be more appropriately called a quantitative approximation of a concept is here called a Platonic Representation by virtue of a vague and misunderstood analogy with Platonic Ideas which were according to the Platonic doctrine, formless and part of his distinction between what is subject to change (to heteron) et what is not affected by change (to auto) [Timeaus, Platon].

I starting to see an increase of random serendipities, perhaps I'm looking for them, in part that's how it works, but...

I was talking to a friend about an idea I had, to build something, he said, "yes! I'm doing that too", then I said "with this specific tooling because is ideal ... " .. he said .. "me too! same tooling", I said "what a coincidence!! amazing ...wait a minute, have you been talking with Claude about this? " ... of course we both did.

My opinion is that that LLMs are influencing us more that we think, and it makes us reach the same conclusions and making us think it was our idea, or that we were the ones steering the conversation.

Perhaps you feel that Claude stole your phrase but I don't think 2 days is enough for that phrase to make it to a re-training, I think it might be more subtle and worse, perhaps you wrote that phrase influenced by the conversation, and the other person did the same.

Just a theory, I can't prove it. But if you think about it, LLMs conversation are the most plausible choices, so it is logical that we all users of this machine are going to tend to converge on the most plausible way of thinking and talking.

No offense, but perhaps that rhyme that you though it was original, it is actually the most plausible in that context.

On the other hand, "Proving and improving" shows a lot of hits in Google Books: https://www.google.com/search?udm=36&q=%22proving+and+improv...

It's possible that you both hit upon or recalled the same cool phrase, especially if the "proving" part was pre-decided and you were thinking of making an incremental improvement.