Summary of Simplicity Bias Via Global Convergence Of Sharpness Minimization, by Khashayar Gatmiry et al.
Simplicity Bias via Global Convergence of Sharpness Minimization
by Khashayar Gatmiry, Zhiyuan Li, Sashank J. Reddi, Stefanie Jegelka
First submitted to arxiv on: 21 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Statistics Theory (math.ST); 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 The study investigates the relationship between the simplicity of neural networks and their generalization ability, particularly in regards to SGD’s implicit bias towards flatter regions of the loss landscape. The authors show that label noise SGD converges to a simple rank one feature matrix for certain activations and small step sizes, which implies a low-rank model. This is achieved by minimizing the sharpness on the manifold of models with zero loss for two-layer networks. The study also discovers a novel property – local geodesic convexity – of the trace of Hessian of the loss at approximate stationary points on this manifold. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research looks into how neural networks get simpler and better at generalizing through something called SGD’s “implicit bias”. It seems that certain ways of training neural networks, like adding noise to the labels, make the models become simpler in a way. The study finds that when using these methods, the model will always end up being very simple and low-rank. This could be important for making AI better at doing things on its own without needing lots of data. |
Keywords
» Artificial intelligence » Generalization
