Summary of An Information-theoretic Analysis Of Compute-optimal Neural Scaling Laws, by Hong Jun Jeon et al.
An Information-Theoretic Analysis of Compute-Optimal Neural Scaling Laws
by Hong Jun Jeon, Benjamin Van Roy
First submitted to arxiv on: 2 Dec 2022
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 The paper studies the optimal trade-off between model size and training data set size for large neural networks. The authors introduce a mathematical framework based on a simple learning model and data generating process, showing that there is a linear relation between these two factors. They derive upper bounds on the minimal achievable expected error as a function of model and data set sizes, and provide empirical results that suggest this approximation correctly identifies an asymptotic linear compute-optimal scaling. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper investigates how to balance the size of neural networks with the amount of training data needed. Researchers have found that using more data doesn’t always help when you’re trying to train a big model. This study looks at why that might be and shows that there’s an ideal balance point where using more data or making the model bigger can improve results. The findings suggest that as you make your models more complex, it makes sense to focus on growing the model rather than collecting more training data. |
