Reproduce “"Understanding Robustness Lottery": A Geometric Visual Comparative Analysis of Neural Network Pruning Approaches” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhimin Li, Shusen Liu, Xin Yu, Kailkhura Bhavya, Jie Cao, Diffenderfer James Daniel, Peer-Timo Bremer, Valerio Pascucci
A visual analytic system is introduced to geometrically compare neural network pruning approaches, revealing how different pruning methods affect internal feature representations, model performance, and robustness. It identifies that Multi-prize Lottery Ticket (MPTs) pruned models possess a latent space geometry less sensitive to data corruptions and discovers redundancies in common evaluation benchmarks.
Key insights
- Multi-prize Lottery Ticket (MPTs) pruned models possess a unique geometric latent space that is inherently less sensitive to data corruptions compared to dense or retrained models.
- Specific geometric properties, such as a large L2 norm and small angle to the true class direction, correlate with a sample's resilience to pruning and corruption.
- Common corruption benchmarks, like Gaussian blur and zoom blur, and clean/brightness/saturation corruptions, induce similar model behaviors, suggesting redundancies in evaluation practices.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.