Summary of Does Sgd Really Happen in Tiny Subspaces?, by Minhak Song et al.
Does SGD really happen in tiny subspaces?
by Minhak Song, Kwangjun Ahn, Chulhee Yun
First submitted to arxiv on: 25 May 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Optimization and Control (math.OC); 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 This paper investigates the training dynamics of deep neural networks by analyzing the relationship between the gradient and the dominant subspace of the loss Hessian during training. The authors find that recent studies’ observation of alignment between the gradient and this subspace is likely spurious, as projecting out the dominant subspace proves to be just as effective as the original update. This phenomenon is observed across various practical setups, including large learning rates, momentum, and adaptive optimizers. The paper discusses the implications of this finding on the dynamics of neural network training. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This study looks at how deep neural networks learn during training. Researchers found that a recent discovery about how gradients relate to the loss Hessian might not be as important as thought. When they tried projecting out some parts of the update, it worked just as well as usual. This happened in different situations, like using big learning rates or special optimizers. The researchers think this will help us understand how neural networks learn better. |
Keywords
» Artificial intelligence » Alignment » Neural network
