Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Transformers Are Graph Neural Networks

thegradient.pub

Post reply on HN
Print view View original

21–26 of 26 posts

Re: Transformers Are Graph Neural Networks

#21

Post by SimplyUnknown »

czzr wrote: post #12
fizixer wrote: post #8

Earlier quoted context omitted.

ML publication is a complete mess right now. Anyone can claim anything as long as they do a write-up and include some equations and pretty plots. It was hard enough 5 years ago to filter out handful of good papers from the sea of bad research. Now it's getting near impossible.

Somebody should train a model to do it.

You mean like arxiv-sanity? As I understand it, it trains a SVM on papers you like to suggest papers that are on the same side of the hyperplane. Could be used as a quality classifier by only liking high-quality work.

Re: Transformers Are Graph Neural Networks

#22

Post by nullc »

fpgaminer wrote: post #9

It's interesting watching these attempts to understand Transformers. Are they Graph networks? Are they Hopfields networks? Are they convolutions? There was this research: https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreti... Turns out, GPT-like architectures appear to use the same representation throughout all the layers. So you can use the final head layer as a lens to see what words the network is thinkin…

With GPT3 you can give it a bit more long term memory by priming it with text that has "self commentary" written and repeated through each paragraph.

[This is a post to Hacker news and I'm making a point to explain a gimmick for giving GPT3 self generated longer term 'memory'.]

Most obvious forms of memory have a problem that they aren't differentiable so you can't train with them in place. This idea works around the issue because english text contains things like running commentary a times, and so a model trained on it already has some idea of how to use it.

[This is a post to Hacker news and I'm making a point to explain a gimmick for giving GPT3 self generated longer term 'memory' and the limitations of other approaches.]

I've had some success at getting this to help generate better text. I wonder though if it would be effective to generate a new training corpus this way. E.g. get GPT3 to generate annotations for arbitrary input text by using some summerization prompt, then use that to go augment the entire training corpus with the summaries injected inline like virtual-thought bubbles with beginning and ending symbols that don't occur in the training material. Then the network is retrained on this augmented data and then can generate its own prompts.

Bonus: the operator could be given access to the otherwise normally hidden "internal monolog" text, to increase control over the output or understand more about the state of the model.

You can't differentiate across the different executions, due to sampling-- but perhaps you don't need to... it doesn't do any gradient descent to perform one shot learning.

I am guessing that this must not work at scale because it's an obvious enough idea and a similar approach for database access (e.g. have it generate keywords from the text, then inject tokens encoding some text search results for those keywords the stream, and skip over them in training and just keep them as context; thus training a model that can use a search to improve its results) must have been tried but I've never heard anyone report it working.

Re: Transformers Are Graph Neural Networks

#23

Post by drran »

fpgaminer wrote: post #9

It's interesting watching these attempts to understand Transformers. Are they Graph networks? Are they Hopfields networks? Are they convolutions? There was this research: https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreti... Turns out, GPT-like architectures appear to use the same representation throughout all the layers. So you can use the final head layer as a lens to see what words the network is thinkin…

Memory is easy. It was proposes 20 years ago, but nobody bothered to translate the paper to English, because 20 years ago AI was a toy.

Memory is just association. When Foo is at input, Memory must bring up Bar, Baz, etc., which are in association with Foo, as separate input. It's better if association kind (before, after, inside, together, opposite, same, etc.) will be stored and retrieved by Memory too. Not a hard task to do by today standards.

However, Long Term Memory is orthogonal to AI training. It's kind of "self-attention" mechanism, because LTM need to watch _training process_, and then note what, when and how put input into LTM, and how to associate it with other things, which are already in LTM. In short, LTM requires meta training, to watch a lot of training sessions, to understand that. It will be hard to define proper loss function for LTM, so it may be better to implement LTM as simple non-AI algorithm first. IMHO, for LTM, rate of training convergence can be used as loss function for meta-training of LTM itself.

BTW, LTM also need a way to translate between input encoding, or single input encoding must be used for all trainings.

PS.

Also, when bringing up associations (memories) for Foo, LTM can also bring up associations for Bar, Baz, etc. For example, LTM can bring up 10 direct (tier 1) associations for Foo, then 3 tier 2 main associations for Bar, Baz, etc., then 1 tier 3 association for tier 2 associations, and so on, up to e.g. 7 tiers. Beware, it can lead to "inner monologue" of machine. :-)

Re: Transformers Are Graph Neural Networks

#24

Post by murbard2 »

fpgaminer wrote: post #9

It's interesting watching these attempts to understand Transformers. Are they Graph networks? Are they Hopfields networks? Are they convolutions? There was this research: https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreti... Turns out, GPT-like architectures appear to use the same representation throughout all the layers. So you can use the final head layer as a lens to see what words the network is thinkin…

> There is absolutely no training signal that tells an RNN to remember something beyond the BPTT horizon. So why would it?

Because they generalize. Char-RNN learn to balance parenthesis separated by a longer distance than the BPTT window because they've learned that counting parenthesis is useful for prediction based on parenthetical statements shorter than the BPTT window.

Re: Transformers Are Graph Neural Networks

#25

Post by unishark »

wnoise wrote: post #7
cjauvin wrote: post #4

Earlier quoted context omitted.

I'm genuinely a bit surprised by that, that was always my high-level understanding of what the essence of neural networks was (at least feedforward vanilla ones), would you care to elaborate?

While they can be thought of as stacked regression, it's only logistic regression with one particular non-linearity. And for many non-linearities you'll have a hard time usefully interpreting them as a regression.

I think the common ones have statistical interpretations (that predate deep learning by a lot). Perhaps the one for the rectified linear unit is pretty obscure. But as I understand it, the statistics concept is called the "Tobit" model. It's meaning is not so obscure though, just a prediction that can be non-negative only, which is pretty common like a mass or energy.

Re: Transformers Are Graph Neural Networks

#26

Post by hrbigelow »

fpgaminer wrote: post #9

It's interesting watching these attempts to understand Transformers. Are they Graph networks? Are they Hopfields networks? Are they convolutions? There was this research: https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreti... Turns out, GPT-like architectures appear to use the same representation throughout all the layers. So you can use the final head layer as a lens to see what words the network is thinkin…

One idea in the Hopfield Networks is All You Need paper, was that the softmax-based attention mechanism is equivalent to a Hopfield energy update, and in which the attention keys are the Hopfield "memories". But, the keys are produced as a transformation of the input, so it seems to me, the Transformer does not actually store keys as "memories" the way a Hopfield network stores memories (as energy minima). Is this correct, or am I missing something about the paper?
Post reply on HN