Summary of Holistic Adversarially Robust Pruning, by Qi Zhao and Christian Wressnegger
Holistic Adversarially Robust Pruning
by Qi Zhao, Christian Wressnegger
First submitted to arxiv on: 19 Dec 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 A novel method called HARP is proposed to compress neural networks while maintaining their accuracy and robustness against adversaries. The approach considers the network holistically, learning a global compression strategy that optimizes which parameters to prune for each layer individually. This is achieved by fine-tuning an existing model with dynamic regularization, balancing different objectives. The learned strategies allow for a significant reduction in network size (99%) while preserving its natural accuracy and adversarial robustness. HARP outperforms prior work in coping with aggressive pruning. The method’s success relies on the consideration of non-uniform compression across layers. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Neural networks can be shrunk to fit small devices, but this often makes them less accurate and vulnerable to attacks. A new way to compress neural networks is proposed, called HARP. Instead of treating each layer separately, HARP looks at the network as a whole. It learns how many parameters to remove and which ones to keep for each layer, based on what works best for that layer. This helps maintain the original accuracy and robustness against attacks. The new method can reduce the size of a network by 99% while keeping it accurate and strong. |
Keywords
» Artificial intelligence » Fine tuning » Pruning » Regularization
