Summary of Efficient Adaptive Optimization Via Subset-norm and Subspace-momentum: Fast, Memory-reduced Training with Convergence Guarantees, by Thien Hang Nguyen et al.
Efficient Adaptive Optimization via Subset-Norm and Subspace-Momentum: Fast, Memory-Reduced Training with Convergence Guarantees
by Thien Hang Nguyen, Huy Le Nguyen
First submitted to arxiv on: 11 Nov 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Neural and Evolutionary Computing (cs.NE); Optimization and Control (math.OC)
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 introduces two efficient adaptive optimization techniques for large-scale neural networks, reducing memory requirements while accelerating training. The first method, Subset-Norm adaptive step size, reduces the memory footprint of the second moment term from O(d) to O(sqrt(d)) through step-size sharing. For non-convex smooth objectives under coordinate-wise sub-gaussian gradient noise, it provides a noise-adapted high-probability convergence guarantee with improved dimensional dependence over existing methods. The second technique, Subspace-Momentum, reduces the momentum state’s memory footprint by operating in a low-dimensional subspace while applying standard SGD in the orthogonal complement. Empirical evaluation on LLaMA models demonstrates the effectiveness of these methods, achieving Adam’s validation perplexity in approximately half the training tokens (6.8B vs 13.1B) with minimal additional hyperparameter tuning. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps us make big neural networks work better and faster. It introduces two new ways to train these networks: one that reduces memory usage and another that uses a special trick to speed things up. Both methods are tested on large language models, showing they can achieve similar results as other methods but using less resources. |
Keywords
» Artificial intelligence » Hyperparameter » Llama » Optimization » Perplexity » Probability
