Reproduce “Sound and Complete Neural Network Repair with Minimality and Locality Guarantees” 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
Feisi Fu, Wenchao Li
REASSURE, a methodology for repairing ReLU neural networks, ensures sound and complete bug correction with theoretical guarantees for locality and minimality in the function space. It consistently outperforms existing methods, achieving significantly lower functional changes (e.g., L∞ ND < 0.002 on MNIST) and preventing negative side effects while preserving original network accuracy.
Key insights
- Leveraging the continuous piecewise linear (CPWL) property of ReLU networks allows for formal guarantees regarding soundness, completeness, locality, and minimality of the repair.
- By constructing a patch network that consists of a localized support function and an affine transformation, REASSURE precisely targets buggy regions in the function space without globally altering the network's behavior.
- Formulating the search for the optimal affine patch function as a robust optimization problem enables efficient computation of minimal functional changes without explicit vertex enumeration in high-dimensional input spaces.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.