Summary of Lotr: Low Tensor Rank Weight Adaptation, by Daniel Bershatsky et al.
LoTR: Low Tensor Rank Weight Adaptation
by Daniel Bershatsky, Daria Cherniuk, Talgat Daulbaev, Aleksandr Mikhalev, Ivan Oseledets
First submitted to arxiv on: 2 Feb 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
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 paper generalizes and extends the idea of low-rank adaptation (LoRA) in large language models (LLMs), specifically Transformer architecture. LoRA-like methods are based on matrix factorization of gradient updates, but this new approach, LoTR, represents updates as tensor decomposition. LoTR constructs a low-rank adapter for each layer as a product of three matrices, sharing left and right multipliers among layers. This allows for simultaneous compression, achieving better parameter efficiency than LoRA, especially in deep models. The core tensor does not depend on original weight dimension, enabling fast and cheap downstream fine-tuning. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps us make big language models work better by finding a new way to update them. Right now, people use something called Low-Rank Adaptation (LoRA) to do this. LoRA is like a special recipe that takes the updates from the model and breaks it down into smaller pieces. This new method, called LoTR, does the same thing but in a more clever way. It uses something called tensor decomposition to update the model’s parameters. This makes the process faster and cheaper, which is great news for people who want to use these models for lots of different tasks. |
Keywords
* Artificial intelligence * Fine tuning * Lora * Low rank adaptation * Transformer
