I'm not as heavy on the maths stuff involved in this as other people commenting appear to be.
But the idea makes sense, of course there is still recoverable data in embeddings, that's the point. Though as I constantly find the more you try to squeeze into an n bit vector the more watered down everything gets.
I suppose a latent space could be encrypted/mapped in some way to resolve that, but how many people are exposing their vectors in the first place?
One way to pose/(think about) the problem is that there are two finite metric spaces linked by an unknown odometry (damn you autocorrect). The problem is to recover that unknown isometry.
This, like graph isometry, can be very computationally intensive in the worst case. However, heuristics to aid matching one vertex on one graph to another vertex on another graph using local, semilocal structural signatures can be very effective on particular cases.
One can of course argue that the spaces are not designed as metric spaces. Even if true, these might be metrizable topological spaces.
More generally, if these are indeed non-metric spaces one can still pose it as finding the unknown isomorphism between two poset spaces.
In my other comment I was using the property of maximal chains -- Identify the longest chains in both posets. The isomorphism must map the longest chain in Poset 1 directly to a longest chain in Poset 2, preserving the exact linear order.
I am not familiar with the standards of publishing in machine learning, but as someone trained in a mathematics background, this paper seems relatively light on details and heavy on exposition. Is that typical? Is this a really novel idea? Not trying to be snarky, just trying to understand how meaningful this is.
It's cool that it proves that a bunch of vectorized outputs from an unknown embedder on an unknown dataset is in no way private, because of this ability to reverse engineer the embedder.
I talked to the author at his poster session at neurips and was able to get the gist, though I had read a lot about the platonic representation hypothesis, and this was one of my top 10 favorite papers in the conference.
They link to their code on GitHub - see footnote 2 on page 2. I don't see it linked anywhere else, which makes it easy to miss. https://github.com/rjha18/vec2vec/
You are not wrong. But this has by no means proven its up to the standard of being publishable in a machine learning journal. Its on arXiv.org, which, lets face it, at the end of the day is a vanity press.
Calling the arxiv a vanity press shows you aren't a researcher. In math and physics all the best stuff is on the arxiv and the general level is well above the level of most journals. Journals mainly serve as accreditation and many are basically mediocre - the review process resta more value than it adds overall.
The pace of things is moving along so rapidly right now, I’m not sure that waiting for peer reviews is always a wise move. Doubly so if there’s a paywall; why limit your article’s impact by placing it where practitioners’ agents might not be able to access it? The rapid progress right now is challenging for conventional academic processes.
If the value of the paper is difficult to independently verify, for example, if it depends on the credibility of the author, then the academic ritual can add something. If it’s a mathematical result, one that can be automatically verified, or a machine learning technique that anyone can try with Claude code reconstructing it for them, this sort of pre-print publishing model is advantageous.
Because...science? It's not science until it passes peer review.
I'm not advocating that everybody stops posting to arXiv, and I'm not saying you can't find good stuff there. I'm just saying, it's a vanity press, there is absolutely no guarantee of the paper's quality.
And being published by a famous professor from a prestigious university is also no guarantee. If we've learned anything from the non-reproducibility crisis, it is that a paper's origin story is no guarantee.
Let's assume that monotonocity of pair-wise distances are preserved.
Without knowing the details of how the paper solved the problem, my first attempt would be to find the diametrically distant pair of points in the two different embeddings and assume that the pair is the same pair. Then find the next distant pairs and so on.
After sufficiently many such pairs have been found, or better still, the largest d-simplex is found, find that scaled rigid body transformation that makes the corresponding pairs coincide. Proceeding this way ought to be less work than solving a generic graph isomorphism problem.
I think a less stringent, but still workable assumption is that for very similair objects, their distances will be small. This is much easier to accomplish than agreement across all pairs.
Could you explain a bit more. What you say about similar objects is obviously true. However the algorithm sketch that you have in your mind is a little implicit. Could you make it more explicit. I am quite curious.
Cyberphrenology. In any two random graphs, you'll find an isomorphic graph which is can be up to log of the size of the graphs.
And if the LLM has been trained up to the limit of what data it can hold, it is going to be random. Proof below if it isn't obvious.
The entire effort of all people who are trying to understand how LLMs work, how they represent their data, its all bound to fail.
Proof: a LLM is a very good approximation of the Solomonov/Levin/Kolmogorov universal probability function on tokens. As such, it will be random--pure white noise--because if you found any patterns in there, you could exploit the regularity and come up with a smaller set of weights for the same LLM.
There are no patterns there to be found. They have all been factored out by training the neural net until it couldn't learn any more.
Distillation is alive and well... Earlier work on model printing also found that it's pretty easy to find smaller sets of parameters which can replicate the behavior of the entire network with pretty good fidelity.
Large parameter counts give space to explore, and give routes out of what would be local minima in a lower dimensional space.
In other words, there's no guarantee that any given trained model is a minimal representation of its training set.
I'm not claiming any arbitrary set of weights is a minimal representation. But typically, if people could achieve the same quality of results with a smaller set of weights, or weights which have been quantized to lower bit representations, etc, they would have published the smaller one instead.
You kind of are claiming they're minimal, though. Because if they're not, your statement that "if you found any patterns in there, you could exploit the regularity..." implies nothing. Yeah, the patterns are there, and people are exploiting them.
Your socioeconomic argument just doesn't hold either. People don't delay releasing models until they've minimized it to the theoretical limit. They ship it when it's good enough for whatever job they're making it for.
If the weights arn't compressed, then a smaller set of weights would perform as well. Sure, you can always induce as much symmetry and patterns as you want by bloating the data set, but that hardly gives us insight into how a set of weights which is "as full as it can be" of information.
The point of TFA is that there are regularities you can exploit in the actually existing weights of machine learning systems, not in some hypothetically maximally efficient weights. The maximally efficient weights would indeed have no structure, but that’s not what anyone is working with.
I'm not as heavy on the maths stuff involved in this as other people commenting appear to be.
But the idea makes sense, of course there is still recoverable data in embeddings, that's the point. Though as I constantly find the more you try to squeeze into an n bit vector the more watered down everything gets.
I suppose a latent space could be encrypted/mapped in some way to resolve that, but how many people are exposing their vectors in the first place?
One way to pose/(think about) the problem is that there are two finite metric spaces linked by an unknown odometry (damn you autocorrect). The problem is to recover that unknown isometry.
This, like graph isometry, can be very computationally intensive in the worst case. However, heuristics to aid matching one vertex on one graph to another vertex on another graph using local, semilocal structural signatures can be very effective on particular cases.
One can of course argue that the spaces are not designed as metric spaces. Even if true, these might be metrizable topological spaces.
More generally, if these are indeed non-metric spaces one can still pose it as finding the unknown isomorphism between two poset spaces.
In my other comment I was using the property of maximal chains -- Identify the longest chains in both posets. The isomorphism must map the longest chain in Poset 1 directly to a longest chain in Poset 2, preserving the exact linear order.
Dupe: https://news.ycombinator.com/item?id=44054425
Note this is version 4 of the paper and the original post was version 1 (I think?)
OpenReview (for NeurIPS) for the curious: https://openreview.net/forum?id=jiCLUPq5xv
I am not familiar with the standards of publishing in machine learning, but as someone trained in a mathematics background, this paper seems relatively light on details and heavy on exposition. Is that typical? Is this a really novel idea? Not trying to be snarky, just trying to understand how meaningful this is.
It's cool that it proves that a bunch of vectorized outputs from an unknown embedder on an unknown dataset is in no way private, because of this ability to reverse engineer the embedder.
I talked to the author at his poster session at neurips and was able to get the gist, though I had read a lot about the platonic representation hypothesis, and this was one of my top 10 favorite papers in the conference.
They link to their code on GitHub - see footnote 2 on page 2. I don't see it linked anywhere else, which makes it easy to miss. https://github.com/rjha18/vec2vec/
It isn't a maths paper, so the conventions are different.
This isn’t math. The exposition IS the details.
You are not wrong. But this has by no means proven its up to the standard of being publishable in a machine learning journal. Its on arXiv.org, which, lets face it, at the end of the day is a vanity press.
Calling the arxiv a vanity press shows you aren't a researcher. In math and physics all the best stuff is on the arxiv and the general level is well above the level of most journals. Journals mainly serve as accreditation and many are basically mediocre - the review process resta more value than it adds overall.
It was accepted to NeurIPS.
The pace of things is moving along so rapidly right now, I’m not sure that waiting for peer reviews is always a wise move. Doubly so if there’s a paywall; why limit your article’s impact by placing it where practitioners’ agents might not be able to access it? The rapid progress right now is challenging for conventional academic processes.
If the value of the paper is difficult to independently verify, for example, if it depends on the credibility of the author, then the academic ritual can add something. If it’s a mathematical result, one that can be automatically verified, or a machine learning technique that anyone can try with Claude code reconstructing it for them, this sort of pre-print publishing model is advantageous.
// why limit your article's impact //
Because...science? It's not science until it passes peer review.
I'm not advocating that everybody stops posting to arXiv, and I'm not saying you can't find good stuff there. I'm just saying, it's a vanity press, there is absolutely no guarantee of the paper's quality.
And being published by a famous professor from a prestigious university is also no guarantee. If we've learned anything from the non-reproducibility crisis, it is that a paper's origin story is no guarantee.
> It's not science until it passes peer review.
You mean Robert Maxwell's quasi-monopoly on scientific publications ?
Then again neither is peer review. Reproducing research is the only way to prove reproducibility and thereby lend credibility to the claims.
At a minimum posting to arxiv gives others a standard way to cite the work.
Let's assume that monotonocity of pair-wise distances are preserved.
Without knowing the details of how the paper solved the problem, my first attempt would be to find the diametrically distant pair of points in the two different embeddings and assume that the pair is the same pair. Then find the next distant pairs and so on.
After sufficiently many such pairs have been found, or better still, the largest d-simplex is found, find that scaled rigid body transformation that makes the corresponding pairs coincide. Proceeding this way ought to be less work than solving a generic graph isomorphism problem.
I think a less stringent, but still workable assumption is that for very similair objects, their distances will be small. This is much easier to accomplish than agreement across all pairs.
Could you explain a bit more. What you say about similar objects is obviously true. However the algorithm sketch that you have in your mind is a little implicit. Could you make it more explicit. I am quite curious.
I explained my thoughts in a comment here
https://news.ycombinator.com/item?id=49595424
I’ve never liked that this was called “the platonic representation hypothesis”. Lots of weird baggage attached and seems like a waste of a good name.
Cyberphrenology. In any two random graphs, you'll find an isomorphic graph which is can be up to log of the size of the graphs.
And if the LLM has been trained up to the limit of what data it can hold, it is going to be random. Proof below if it isn't obvious.
The entire effort of all people who are trying to understand how LLMs work, how they represent their data, its all bound to fail.
Proof: a LLM is a very good approximation of the Solomonov/Levin/Kolmogorov universal probability function on tokens. As such, it will be random--pure white noise--because if you found any patterns in there, you could exploit the regularity and come up with a smaller set of weights for the same LLM.
There are no patterns there to be found. They have all been factored out by training the neural net until it couldn't learn any more.
/a smaller set of weights for the same LLM./
Distillation is alive and well... Earlier work on model printing also found that it's pretty easy to find smaller sets of parameters which can replicate the behavior of the entire network with pretty good fidelity.
Large parameter counts give space to explore, and give routes out of what would be local minima in a lower dimensional space.
In other words, there's no guarantee that any given trained model is a minimal representation of its training set.
I'm not claiming any arbitrary set of weights is a minimal representation. But typically, if people could achieve the same quality of results with a smaller set of weights, or weights which have been quantized to lower bit representations, etc, they would have published the smaller one instead.
You kind of are claiming they're minimal, though. Because if they're not, your statement that "if you found any patterns in there, you could exploit the regularity..." implies nothing. Yeah, the patterns are there, and people are exploiting them.
Your socioeconomic argument just doesn't hold either. People don't delay releasing models until they've minimized it to the theoretical limit. They ship it when it's good enough for whatever job they're making it for.
The weights aren’t compressed. So there are interpretable redundancies in practice.
If the weights arn't compressed, then a smaller set of weights would perform as well. Sure, you can always induce as much symmetry and patterns as you want by bloating the data set, but that hardly gives us insight into how a set of weights which is "as full as it can be" of information.
The point of TFA is that there are regularities you can exploit in the actually existing weights of machine learning systems, not in some hypothetically maximally efficient weights. The maximally efficient weights would indeed have no structure, but that’s not what anyone is working with.
What is the (co)homology of this space?
That of the underlying, hypothetical universal brain topology?