Summary of A Method on Searching Better Activation Functions, by Haoyuan Sun et al.
A Method on Searching Better Activation Functions
by Haoyuan Sun, Zihao Wu, Bo Xia, Pu Chang, Zibin Dong, Yifu Yuan, Yongzhe Chang, Xueqian Wang
First submitted to arxiv on: 19 May 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 paper presents a theoretical and experimental approach to optimizing activation functions in artificial neural networks (ANNs). The authors demonstrate the existence of the worst activation function with boundary conditions (WAFBC) from an information entropy perspective. They propose a novel methodology, Entropy-based Activation Function Optimization (EAFO), which enables the design of static activation functions and dynamic optimization during iterative training. The authors derive a new activation function, Correction Regularized ReLU (CRReLU), using EAFO. Experiments on vision transformer variants show CRReLU outperforms existing corrections of ReLU on CIFAR-10, CIFAR-100, and ImageNet-1K datasets. Furthermore, CRReLU exhibits superior performance compared to GELU in large language model fine-tuning tasks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper tries to find the best way to make artificial neural networks work better by choosing the right “glorified math” (activation function) that helps them learn. They look at how much information is lost when you choose a bad activation function and come up with a new way to design good ones. This new approach, called EAFO, lets them make static activation functions and even change them during training. The authors then create a new type of activation function, CRReLU, which beats other versions on some big datasets. |
Keywords
» Artificial intelligence » Fine tuning » Large language model » Optimization » Relu » Vision transformer
