Summary of Lory: Fully Differentiable Mixture-of-experts For Autoregressive Language Model Pre-training, by Zexuan Zhong et al.
Lory: Fully Differentiable Mixture-of-Experts for Autoregressive Language Model Pre-training
by Zexuan Zhong, Mengzhou Xia, Danqi Chen, Mike Lewis
First submitted to arxiv on: 6 May 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: 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 proposed Lory approach is a scalable, fully-differentiable Mixture-of-Experts (MoE) architecture that efficiently trains autoregressive language models. This approach introduces two key techniques: causal segment routing for efficient expert merging operations and similarity-based data batching to encourage expert specialization. Lory models are pre-trained on 150B tokens from scratch with up to 32 experts and 30B parameters, achieving significant performance gains over dense models in perplexity (+13.9%) and downstream tasks (+1.5%-11.1%). The trained experts capture domain-level specialization without supervision. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Lory is a new way to train language models that helps them learn from lots of text data more efficiently. It’s like having many small helpers, each one specialized in a certain type of writing or speaking. This approach makes the training process faster and better, and it can be used for many different tasks, such as understanding spoken language or generating human-like text. |
Keywords
» Artificial intelligence » Autoregressive » Mixture of experts » Perplexity
