Summary of Achieving Sparse Activation in Small Language Models, by Jifeng Song et al.
Achieving Sparse Activation in Small Language Models
by Jifeng Song, Kai Huang, Xiangyu Yin, Boyuan Yang, Wei Gao
First submitted to arxiv on: 3 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 A novel approach to sparse activation in Small Language Models (SLMs) is proposed, aiming to reduce the computing cost of SLMs without retraining or adaptation efforts. Building on neurons’ output magnitudes is not suitable for SLMs due to their lower over-parameterization compared to Large Language Models (LLMs). Instead, activating neurons based on attribution scores proves more effective. However, existing attribution metrics are inaccurate when used for sparse activation due to interdependency among scores across layers. To address this, a new attribution metric is proposed, which can correct errors and achieve precise sparse activation. Experimental results show that the approach achieves an 80% sparsification ratio with less than 5% model accuracy loss, comparable to LLMs. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Sparse activation in Small Language Models (SLMs) is explored to reduce computing costs without retraining. Existing methods based on neurons’ output magnitudes don’t work well for SLMs because they’re not over-parameterized like Large Language Models (LLMs). Instead, using attribution scores helps. But existing metrics are bad at sparse activation due to how they’re connected across layers. A new metric is developed to fix this and get precise activation. Tests show that it can achieve 80% sparsification with only a small loss in accuracy. |
