Maximal Robust Neural Network Specifications via Oracle-guided Numerical Optimi...
summary
MaRVeL computes maximal non-uniform robust specifications that explicitly maximize a chosen norm by coupling a numerical optimizer with a non-differentiable verifier as an oracle-guided numerical optimization loop. The optimizer proposes candidate specifications, the verifier returns closest inputs on decision boundaries, and the optimizer uses those points to compute gradients and expand the specification while preserving robustness. Evaluation across datasets and classifiers yields specifications 5.1x larger than prior methods and on a 2D dataset reaches 93% of optimal average diameter versus 26% for earlier work.
AI-written summary · about
related talks
01h 23mImplicit Regularization II
16mVerification of Neural Networks' Global Robustness
01h 17mRobust Optimization and Generalization
01h 01mRobustness, Stochastics, Uncertainty 2
49mOracle-Efficient Online Learning or: How to Use Non-Robust Optimization for Robust Learning
01h 01mML for discrete optimization: Theoretical foundations
46mSmooth Nonconvex Min-Max Optimization
46mCan Non-Convex Optimization be Robust?