Thread by @gordic_aleksa

[🧠 Paper Summary 📚] An interesting paper was recently published to arxiv: "Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets" (although it originally appeared in May 2021).

The main idea is this:
1/ 🧵

If you have an overparametrized neural network (more params than the # of data points in your dataset) and you train it way past the point where it has memorized the training data (as suggested by the low training loss, and high val loss), all of a sudden the network will...

2/

learn to generalize, as suggested by a rapid decrease in the val loss (colloquially known as "grok" i.e. the NN understood/figured it out).

Practitioners usually stop training networks at the 1st sign of overfitting (as evidenced by an increasing gap between train/val loss).

3/

Interestingly enough, the authors of the paper forgot to turn off the model training and that led to this paper haha! @OpenAI @exteriorpower

The phenomenon is particularly evident on smaller datasets.

This goes against the common understanding in statistics that suggests...

4/

that you want to have underparametrized models so that you force the model to learn the rule (and thus generalize to unseen data) instead of memorizing the training dataset.

The trick is that heavily overparametrized models combined with especially weight decay seem to be...

5/

able to "carve out" that simple model from the big overparametrized NN.

Also, again, there seems to be some correlation between flatter loss regions and good generalization capabilities. It may be that weight decay is especially good at leading the optimization process...

6/

towards those flatter regions of the loss landscape.

Check out the paper here:
arxiv.org/abs/2201.02177

Also, @ykilcher nicely explained it here:


#deeplearning #overparametrization #grokking

7/7