Transformers Are Graph Neural Networks | Hacker News Reader

Transformers Are Graph Neural Networks

graphdeeplearning.github.io

142 pointsAnon8419 comments

The top level site is pretty neat [1]! I've been trying to find a site like this to learn more about the applications of graph NN. I particularly like the one about combinatorial optimization/discrete optimization.

[1] https://graphdeeplearning.github.io/

You might also be interested in the project-based course on GNNs by octavian.ai which is here: https://www.octavian.ai/machine-learning-on-graphs-course You don't need to sign up -- the notebooks get posted to this page. Signup is only if you want to receive notifications.

Thanks for highlighting! (Admin of the lab website)

> Reading new Transformer papers makes me feel that training these models requires something akin to black magic when determining the best learning rate schedule, warmup strategy and decay settings.

This is the slow decline of the machine learning field because most researchers are too busy procuring positions in various institutions instead of asking and pursuing more creative and more difficult lines of question/thought.

I am glad the author chooses to call out disconcerting behavior.

> these papers showed that Transformer heads can be ‘pruned’ or removed after training without significant performance impact.

The larger the model the less important or just purely redundant the various modules will become.

The ML community has seen spectacularly productive in the last few years, comparable only to the inventors of dark mode. The results frequently capture the public imagination. For every brogrammer griping about how “it’s all basic statistics” there are hundreds of people thinking magic must be involved.

Correlation may be != causation. But absent other evidence, I would still be careful with accusations of any systemic issues with their work culture, or predictions about impending doom.

I seem to remember some of these seemingly secondary parameters at one point being the sole reason making the model work. Wasn’t it a new initialization schedule that kicked of the current boom?

In any case, recent history should be a good example of how “just” uncreative pursuits such as increasing model depth can have results of dramatically different quality.

It also feels strange to take issue with people being motivated by publication. As far as inducing altruistic behavior goes, publications are second only to cheap medals handed posthumously to the children of dead soldiers. And in terms publication criteria being aligned with some abstract sense of “good research”, I have little doubt that creative ideas with some interesting results will find an interested audience. There may well remain the universal problem that unsuccessful “out-there” efforts may leave you with little when they fail, but risk is as inherent to such efforts as the chance to make it big. It’s almost tautologically impossible to adequately reward failed efforts, because there are no measures to asses them; indeed, where there are measure, they are no longer deemed to have failed.

To then make it easier to advance along new lines of thinking, we would want to come up with new yardsticks to judge results: coming up with new standard problem sets where current efforts fail dramatically. Thinking back over the last years, I’m not entirely sure that isn’t exactly what we’ve been doing.

>> Correlation may be != causation. But absent other evidence, I would still be careful with accusations of any systemic issues with their work culture, or predictions about impending doom.

No lesser man than Geoff Hinton himself thinks there are systemic issues with machine learning publications, although he doesn't foresee impending doom:

WIRED: The recent boom of interest and investment in AI and machine learning means there’s more funding for research than ever. Does the rapid growth of the field also bring new challenges?

GH: One big challenge the community faces is that if you want to get a paper published in machine learning now it's got to have a table in it, with all these different data sets across the top, and all these different methods along the side, and your method has to look like the best one. If it doesn’t look like that, it’s hard to get published. I don't think that's encouraging people to think about radically new ideas.

Now if you send in a paper that has a radically new idea, there's no chance in hell it will get accepted, because it's going to get some junior reviewer who doesn't understand it. Or it’s going to get a senior reviewer who's trying to review too many papers and doesn't understand it first time round and assumes it must be nonsense. Anything that makes the brain hurt is not going to get accepted. And I think that's really bad.

What we should be going for, particularly in the basic science conferences, is radically new ideas. Because we know a radically new idea in the long run is going to be much more influential than a tiny improvement. That's I think the main downside of the fact that we've got this inversion now, where you've got a few senior guys and a gazillion young guys.

https://www.wired.com/story/googles-ai-guru-computers-think-...

We are pushing boundaries in the industry, sometimes in secret, but often cases making public contributions that are beyond the resources of the private sector.

That's actually been part of the reason for it's explosion. See Google, Facebook, Amazon, Uber, all releasing public blurbs and repos of their research. And papers.

The multidisciplinary nature of ML attracts science types who value publications. It's a sort of club. I'm talking about the people doing really cool and difficult stuff - robots, AI, self driving, cutting edge and high latent value. It's a strong incentive for them.

Yeah, I guess this is fine if by "Transformers are Graph Neural Networks" we mean Transformers < GNN, rather than Tranformers == GNN.

"Sentences are a fully connected graph". Ok fine, but that's a graph with basically no information embedded in its structure. GNNs are supposed to be useful for graphs that have interesting structure, right?

Not necessarily. In fact, applying the Transformer/GNN on a full graph is seen by some people as 'discovering' or identifying some useful underlying graph structure. We had an interesting discussion on this on Twitter: https://twitter.com/PetarV_93/status/1233820829852618754

It says a lot about my knowledge of ML that I went into this post fully expecting it be about Autobots and Decepticons.

Does the use of positional embeddings mess with the gnn for formulation? I'm not familiar with the requirements for something to be a gnn, but positional embeddings mean the graphs have to capture order of occurrence, and the graph in the shared page doesn't seem to do that.

In graph terms, Positional encodings are useful for adding sequential/temporal properties to each node in the graph. Indeed, there are works on position-aware GNNs.

Sorry but I hate it. Graph->graph transform can describe almost any operation, it's a superset of linear transforms. When all you know about your model is that 'it's a graph', you barely know anything at all. And transformers are just glorified AST parsers.

> And transformers are just glorified AST parsers.

Can you explain this?

I am guessing he means that Transformers 'discover' parse trees in sentences, as they operate on fully-connected graphs but end up focusing on the most important connections through attention.

Graph attention networks were inspired by transformers. I am pretty sure they explicitly say it in the paper.

Indeed, the two papers came out within months of each other iirc. The GAT paper discusses Transformers in the context of stabilizing the learning of attention mechanisms.

Of course, this connection may be trivial to most people, but I hadn't seen a post on this before. So I decided to write one for myself as I studied these architectures.

Maybe a better way to put it would be "Graph Neural Networks are a Generalization of Transformers". The formulation of "X are Y" indeed suggests that you could just have stuck with Y instead of using the special case X, whereas in reality they took Y and added something to make it more general.

Yeah but even without attention, they're pretty similar.

Sentences can be reasonable modeled as a mostly 1 directional graph.

Reply on news.ycombinator.com