Summary of Lolcats: on Low-rank Linearizing Of Large Language Models, by Michael Zhang et al.
LoLCATs: On Low-Rank Linearizing of Large Language Models
by Michael Zhang, Simran Arora, Rahul Chalamala, Alan Wu, Benjamin Spector, Aaryan Singhal, Krithik Ramesh, Christopher Ré
First submitted to arxiv on: 14 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); 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 Recent advancements in large language models (LLMs) have shown that linearizing quadratic attentions can significantly reduce pretraining costs. However, this often comes at the cost of decreased model quality and limited scalability to larger LLMs. To address these limitations, we propose LoLCATs, a simple two-step method for improving LLM linearization quality while reducing memory and compute requirements by orders of magnitude. This approach is based on two key findings: replacing softmax attentions with linear attentions through attention transfer, and adjusting for approximation errors using low-rank adaptation (LoRA). Our results demonstrate that LoLCATs significantly improves linearizing quality, training efficiency, and scalability, reducing the gap between linearized and original LLMs by up to 78.1% on the MMLU benchmark. Additionally, we apply LoLCATs to create the first linearized 70B and 405B LLMs, achieving state-of-the-art performance. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Scientists have been working on a way to make large language models (LLMs) more efficient without sacrificing their ability to understand and generate text. They’ve found that by “linearizing” the model’s attention mechanism, they can reduce the amount of computing power needed to train it. However, this often comes at a cost: the linearized model isn’t as good as the original one. To fix this problem, researchers have developed a new method called LoLCATs that improves the quality of the linearized model while reducing the need for computing resources. This means they can create larger and more powerful LLMs without breaking the bank or requiring too much processing power. |
Keywords
» Artificial intelligence » Attention » Lora » Low rank adaptation » Pretraining » Softmax
