Summary of Understanding and Minimising Outlier Features in Neural Network Training, by Bobby He et al.
Understanding and Minimising Outlier Features in Neural Network Training
by Bobby He, Lorenzo Noci, Daniele Paliotta, Imanol Schlag, Thomas Hofmann
First submitted to arxiv on: 29 May 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
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 investigates the phenomenon of Outlier Features (OFs) in neural networks, which are neurons whose activation magnitudes significantly exceed the average over a network’s width. OFs have been observed to emerge during standard transformer training and negatively impact quantization in affected models. Despite their practical importance, little is known about why OFs arise during training or how they can be minimized. The paper aims to shed light on these questions by exploring the underlying mechanisms driving OF emergence and proposing methods for reducing their occurrence. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research looks at something called Outlier Features in computer networks. These features are special neurons that get really strong signals compared to others, which is bad because it makes it hard to make the network smaller and more efficient. Nobody knows why these features show up or how to stop them from happening, so this paper tries to figure out what’s going on and find ways to fix the problem. |
Keywords
» Artificial intelligence » Quantization » Transformer
