Summary of Sltrain: a Sparse Plus Low-rank Approach For Parameter and Memory Efficient Pretraining, by Andi Han et al.
SLTrain: a sparse plus low-rank approach for parameter and memory efficient pretraining
by Andi Han, Jiaxiang Li, Wei Huang, Mingyi Hong, Akiko Takeda, Pratik Jawanpuria, Bamdev Mishra
First submitted to arxiv on: 4 Jun 2024
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 A novel approach for efficient pretraining of large language models (LLMs) is proposed in this work. The authors introduce a method called SLTrain, which combines low-rank and sparse matrix factorization to reduce computational requirements while maintaining performance. By parameterizing weights as a sum of these two components, the model learns to represent knowledge more efficiently. Experimental results demonstrate that SLTrain achieves comparable performance to full-rank training with minimal additional parameters and memory costs. When combined with quantization and per-layer updates, SLTrain can further reduce memory requirements by up to 73% when pretraining the LLaMA 7B model. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models (LLMs) are really smart computers that can understand and generate human-like text. But they require a lot of computing power and memory to train. Some researchers have found ways to make training more efficient by using special tricks on the weights. This paper proposes a new trick called SLTrain, which combines two ideas: reducing the number of unique values in the weights (low-rank) and making some parts of the weights zero (sparse). The results show that this method is very effective at training LLMs without needing too much computing power or memory. |
Keywords
» Artificial intelligence » Llama » Pretraining » Quantization
