To me, the main insight in the paper is that the norm of the initial weights matter, and not just weight decay/other forms of regularization! IE, while people have plotted weight norm as a function of training time for grokking/nongrokking networks, people have not, afaik, plotted initial weight norm vs grokking.  I had originally thought that overfitting would just happen, since memorization was "easier" in some sense for SGD to find in general. So it's a big update for me that the initial weight norm matters so much. 

I don't think grokking per se is particularly important, except insofar as neat puzzles about neural networks are helpful for understanding neural networks. As both this paper and Neel/Lieberum's grokking post argue, grokking happens when there's limited data that's trained for a long time, which causes SGD to initially favor memorization, and some form of regularization, which causes the network to eventually get to a generalizing solution. (Assuming large enough initializations.) But in practice, large foundation models are not trained for tens of thousands of epochs on tiny datasets, but instead a single digit number of epochs on a large dataset (generally 1).  Also, if the results from this paper (where you need progressively larger weight initializations to get grokking with larger models/larger datasets, then it seems unlikely that any large model is in the grokking regime). 

So I think there's something else going on behind the rapid capability gains we see in other networks as we scale the amount of training data/train steps/network parameters. And I don't expect that further constraining the weight norm will speed up generalization on current large models. 

That being said, I do think the insight that the norms of the initial weight matter for generalization seems pretty interesting!

Nitpick: I'm really not a fan of people putting the number of steps on a log scale, since it makes grokking look far more sudden than it actually is, while making the norm's evolution look smoother than it is. Here's what the figures look like for 5 random seeds on the P=113 modular addition task, if we don't take log of the x axis: https://imgur.com/a/xNhHDmR
 
Another nitpick: I thought it was confusing that the authors used "grokking" to mean "delay in generalization" and "de-grokking" to mean "generalization". This seems the opposite of what "grok" actually means?
 

EDIT: Also, I'm not sure that I fully understand or buy the claim that "representation learning is key to grokking".

Hmm, I haven't read the paper yet, but thinking about it, the easiest way to change the weight norm is just to multiply all weights by some factor $\alpha$, but then in a network with ReLU activations and L layers, this would be completely equivalent to multiplying the output of the network by ${\alpha }^L$. In the overfitting regime where the network produces probability distributions where the mode is equal to the answer for all training tokens, the easiest way to decrease loss is just to multiply the output by a constant factor, essentially decreasing the entropy of the distribution forever, but this strategy fails at test-time because the mode is not equal to the answer there. So keeping the weight norm at some specified level might just be a way to prevent the network from taking the easy way towards decreasing training loss, and forcing it to find ways at constant-weight-norm to decrease loss, which would better generalize for the test-set.

One of the authors of the paper here. Glad you found it interesting! In case people want to mess around with some of our results themselves, here are colab notebooks for reproducing a couple results:

1. Delaying generalization (inducing grokking) on MNIST: https://colab.research.google.com/drive/1wLkyHadyWiZSwaR0skJ7NypiYKCiM7CR?usp=sharing
2. Almost eliminating grokking (bringing train and test curves together) in transformers trained on modular addition: https://colab.research.google.com/drive/1NsoM0gao97jqt0gN64KCsomsPoqNlAi4?usp=sharing

Some miscellaneous comments:

• On some level "just fix your weight norm and the model generalizes" sounds too simple to be true for all tasks -- I agree. I'd be pretty surprised if our result on speeding up generalization on modular arithmetic by constraining weight norm had much relevance to training large language models, for instance. But I haven't thought much about this yet!
• In terms of relevance to AI safety, I view this work broadly as contributing to a scientific understanding of emergence in ML c.f. "More is Different for AI". It seems useful for us to understand mechanistically how/why surprising capabilities are gained in increasing model scale or training time (as is the case for grokking), so that we can better reason about and anticipate the potential capabilities and risks of future AI systems. Another AI safety angle could lie in trying to unify our observations with Nanda and Lieberum's circuits-based perspective on grokking. My understanding of that work is that networks learn both memorizing and generalizing circuits, and that generalization corresponds to the network eventually "cleaning up" the memorizing circuit, leaving the generalizing circuits. By constraining weight norm, are we just preventing the memorizing circuits from forming? If so, can we learn something about circuits, or auto-discover them, by looking at properties of the loss landscape? In our setup, does switching to polar coordinates factor the parameter space into things which generalize and things which memorize, with the radial direction corresponding to memorization and the angular directions corresponding to generalization? Maybe there are general lessons here.
• Razied's comment makes a good point about weight L2 norm being a bizarre metric for generalization, since you can take a ReLU network which generalizes and arbitrarily increase its weight norm by multiplying neuron in-weights by $\beta$ and its out-weights by $1/\beta$ without changing the function implemented by the network. The relationship between weight norm and generalization is an imperfect one. What we find empirically is simply this: when we initialize networks in a standard way, multiply all the parameters by $\alpha$, and then constrain optimization to lie on that constant-norm sphere in parameter space, there is often an $\alpha$-dependent gap in test and train performance for the solutions that optimizers find. For large $\alpha$, optimization finds a solution on the sphere which fits the training data but doesn't generalize. For $\alpha$ in the right range, optimization finds a solution on the sphere which does generalize. So maybe the right statement about generalization and weight norm is more about the density of generalizing vs not generalizing solutions in different regions of parameter space, rather than their existence. I'll also point out that this gap between train and test performance as a function of $\alpha$ is often only present when we reduce the size of the training dataset. I don't yet understand mechanistically why this last part is true.