Summary of Evolving Subnetwork Training For Large Language Models, by Hanqi Li et al.
Evolving Subnetwork Training for Large Language Models
by Hanqi Li, Lu Chen, Da Ma, Zijian Wu, Su Zhu, Kai Yu
First submitted to arxiv on: 11 Jun 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
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 proposes a novel training paradigm called Evolving Subnetwork Training (EST) to reduce the training costs of large language models. EST samples subnetworks from the layers and modules of existing models, such as GPT2 and TinyLlama, and gradually increases their size during training, achieving a 26.7% FLOPs saving for GPT2 and 25.0% for TinyLlama without sacrificing performance on pre-training datasets. The approach also leads to performance improvements in downstream tasks, indicating better generalization. Theoretical studies based on training dynamics and Dropout theory ensure the feasibility of EST. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models have revolutionized AI research, but their high training costs hinder further development and adoption. A new approach called Evolving Subnetwork Training (EST) can help reduce these costs while maintaining performance. EST works by selecting small parts of existing large models, growing them during training, and saving the cost of training. This approach was tested on two models, GPT2 and TinyLlama, and showed a 26.7% and 25.0% reduction in computing power needed for training without losing accuracy. Additionally, EST improved performance in tasks that use these models. |
Keywords
» Artificial intelligence » Dropout » Generalization
