Summary of Gw-moe: Resolving Uncertainty in Moe Router with Global Workspace Theory, by Haoze Wu et al.
GW-MoE: Resolving Uncertainty in MoE Router with Global Workspace Theory
by Haoze Wu, Zihan Qiu, Zili Wang, Hang Zhao, Jie Fu
First submitted to arxiv on: 18 Jun 2024
Categories
- Main: Machine Learning (cs.LG)
- 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 proposed Global Workspace-Mixture-of-Experts (GW-MoE) fine-tuning method addresses the issue of uncertain routing results in Mixture-of-Experts models. MoE has been shown to be an efficient way to scale up models by dynamically selecting activated experts, but this approach can lead to incorrect selections when tokens have nearly equal scores for choosing each expert. GW-MoE broadcasts uncertain tokens across experts during fine-tuning, allowing these tokens to acquire necessary knowledge from any expert and become less sensitive to choice. This method does not introduce additional inference overhead and consistently improves performance in various tasks (text classification, question answering, summarization, code generation, and mathematical problem solving) and model sizes (650M and 8B parameters). |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary A new way to improve Mixture-of-Experts models is proposed, called Global Workspace-MoE. The current method has a problem when some tokens have similar scores for each expert, which can lead to wrong choices. To fix this issue, the new method shares information about these uncertain tokens among all experts during training. This helps these tokens learn from any expert and make better decisions. This new approach doesn’t slow down the model’s predictions and works well in different tasks like classifying text, answering questions, summarizing text, generating code, and solving math problems. |
Keywords
» Artificial intelligence » Fine tuning » Inference » Mixture of experts » Question answering » Summarization » Text classification
