Summary of Deep Grokking: Would Deep Neural Networks Generalize Better?, by Simin Fan et al.
Deep Grokking: Would Deep Neural Networks Generalize Better?
by Simin Fan, Razvan Pascanu, Martin Jaggi
First submitted to arxiv on: 29 May 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Machine Learning (stat.ML)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Deep learning researchers have long been fascinated by the “grokking” phenomenon, where neural networks exhibit a sudden surge in test accuracy long after overfitting. This phenomenon has primarily been studied on shallow networks, but our paper explores its occurrence on deeper networks like 12-layer MLPs. We replicate the phenomenon and find that deep networks can be more susceptible to grokking than shallower ones. Our results also show a multi-stage generalization pattern in test accuracy, which is less common in shallow models. We investigate feature ranks and their relationship with generalization performance, suggesting that internal feature rank could be a better indicator of model behavior than weight norms. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Deep learning is like trying to figure out a puzzle. Researchers have discovered something called “grokking” where the puzzle pieces (called neural networks) fit together perfectly at first, but then suddenly start fitting into a bigger picture (generalizing well). This usually happens after the network has learned everything about the training data. Our study looks at what happens when we make the neural network deeper, like having more layers. We found that deeper networks can be even better at grokking than shallower ones. We also saw that the puzzle pieces start to fit together in a different way as the network gets deeper, which is unusual. This tells us that there’s something special about how the features (or puzzle pieces) are arranged inside the network that helps it generalize well. |
Keywords
» Artificial intelligence » Deep learning » Generalization » Neural network » Overfitting
