Summary of Exploring the Benefit Of Activation Sparsity in Pre-training, by Zhengyan Zhang et al.
Exploring the Benefit of Activation Sparsity in Pre-training
by Zhengyan Zhang, Chaojun Xiao, Qiujieli Qin, Yankai Lin, Zhiyuan Zeng, Xu Han, Zhiyuan Liu, Ruobing Xie, Maosong Sun, Jie Zhou
First submitted to arxiv on: 4 Oct 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 This paper explores the potential of pre-trained Transformers by leveraging their inherent sparse activation property. The authors study how activation properties change during pre-training and find that Transformers exhibit sparse activation throughout most of the process, with an evolving correlation between activations as training progresses. To capitalize on this finding, the authors propose Switchable Sparse-Dense Learning (SSD), which adaptively switches between Mixtures-of-Experts (MoE) based sparse training and conventional dense training during pre-training. SSD achieves comparable performance to dense training with reduced costs and can be used as MoE models for sparse inference, achieving up to 2x faster inference speed compared to dense models. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how transformers learn from data before being trained on a specific task. The authors found that most of the time, only some parts of the transformer are “turned on” and working together. They used this finding to create a new way of training called SSD (Switchable Sparse-Dense Learning). This method helps transformers learn faster and use less energy during the learning process. The results show that using SSD can help achieve similar performance as traditional methods, but with fewer resources needed. |
Keywords
» Artificial intelligence » Inference » Transformer
