Summary of How Does Overparameterization Affect Features?, by Ahmet Cagri Duzgun et al.
How Does Overparameterization Affect Features?
by Ahmet Cagri Duzgun, Samy Jelassi, Yuanzhi Li
First submitted to arxiv on: 1 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
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 This paper investigates the characteristics of features learned by overparameterized deep learning models, which have more parameters than necessary to fit their training loss. The authors compare models with the same architecture but different widths and find that both overparameterized and underparameterized networks acquire unique features. They show that the feature space of overparameterized networks cannot be spanned by concatenating many underparameterized features, and vice versa. The results suggest that overparameterized networks outperform underparameterized networks, even when many of the latter are concatenated. This is demonstrated using a VGG-16, ResNet18 on CIFAR-10, and a Transformer on the MNLI classification dataset. The authors also propose a toy setting to explain how overparameterized networks can learn important features that underparamaterized networks cannot. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how deep learning models work when they have more parts than they need to learn. The researchers compared different versions of the same model and found that each one is good at learning certain things. They showed that these models are really good at finding patterns in data, even when we combine many simpler models together. This means that using a lot of extra information can actually make our models worse! The results were tested on lots of different datasets and show that more complex models can be better than simpler ones. |
Keywords
* Artificial intelligence * Classification * Deep learning * Transformer
