Summary of Revealing the Utilized Rank Of Subspaces Of Learning in Neural Networks, by Isha Garg et al.
Revealing the Utilized Rank of Subspaces of Learning in Neural Networks
by Isha Garg, Christian Koguchi, Eshan Verma, Daniel Ulbricht
First submitted to arxiv on: 5 Jul 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- 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 work examines how well the learned weights of a neural network utilize the available space, considering both capacity and interactions with the dataset. The findings suggest that most learned weights are full rank, implying they use the entire space available to them. However, a simple data-driven transformation can project these weights onto the subspace where the data and weight interact, revealing low-rank structure while preserving functional mapping. This leads to reducing parameters by up to 75% with minimal accuracy drops (less than 0.2%) after fine-tuning. The study also highlights that self-supervised pre-training drives utilization rates up to 70%, justifying its suitability for downstream tasks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper looks at how well neural networks use the space available to them. It seems like most of the network’s weights are using all the space they have, but a simple trick can help us see that this isn’t actually true. By projecting the weights onto a specific subspace, we can find out that many models only use about 20-35% of their available space! This is important because it means we can make these models smaller and more efficient without losing much accuracy. The study also shows that using self-supervised learning to pre-train the networks helps them use even less space while still performing well. |
Keywords
* Artificial intelligence * Fine tuning * Neural network * Self supervised
