Summary of How Feature Learning Can Improve Neural Scaling Laws, by Blake Bordelon et al.
How Feature Learning Can Improve Neural Scaling Laws
by Blake Bordelon, Alexander Atanasov, Cengiz Pehlevan
First submitted to arxiv on: 26 Sep 2024
Categories
- Main: Machine Learning (stat.ML)
- Secondary: Disordered Systems and Neural Networks (cond-mat.dis-nn); Machine Learning (cs.LG)
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 proposes a novel solvable model of neural scaling laws, examining how performance changes with model size, training time, and data availability. Theoretical analysis identifies three regimes: hard, easy, and super-easy tasks, which exhibit distinct scaling exponents. For easy and super-easy tasks within the reproducing kernel Hilbert space (RKHS), scaling exponents remain consistent between feature learning and kernel regime models. In contrast, for hard tasks outside the RKHS, feature learning improves scaling with training time and compute, nearly doubling the exponent. This leads to a different optimal strategy for scaling parameters and training time in the feature learning regime. The study supports its findings through experiments on nonlinear MLPs fitting functions with power-law Fourier spectra and CNNs learning vision tasks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper looks at how well AI models perform as they get bigger and train more. They found that the way performance changes depends on the difficulty of the task. For easy tasks, the model’s size doesn’t matter much, but for harder tasks, having a larger model can really help. The study also showed that when learning features from data, the model can improve its ability to perform well with harder tasks. |
Keywords
» Artificial intelligence » Scaling laws
