Sorting out Lipschitz function approximation
Cem Anil, James Lucas, Roger Grosse · arXiv (Cornell University) · 2018
Training neural networks under a strict Lipschitz constraint is useful for provable adversarial robustness, generalization bounds, interpretable gradients, and Wasserstein distance estimation. By the composition property of Lipschitz functions, it suffices to ensure that each individual affine transformation or nonlinear activation is 1-Lipschitz. The challenge is to do this while maintaining the expressive power. We identify a necessary property for such an architecture: each of the layers must preserve the gradient norm during backpropagation. Based on this, we propose to combine a gradient norm preserving activation function, GroupSort, with norm-constrained weight matrices. We show that norm-constrained GroupSort architectures are universal Lipschitz function approximators. Empirically, we show that norm-constrained GroupSort networks achieve tighter estimates of Wasserstein distance than their ReLU counterparts and can achieve provable adversarial robustness guarantees with little cost to accuracy.
Cited by 58 works. doi:10.48550/arxiv.1811.05381 Open-access PDF
Authors
References
- Deep Residual Learning for Image Recognition
- AutoResearch Lite: a reproducible paper-inspired experiment runner (Adam mechanism reproduction and convergence-speed comparison)
- Learning Multiple Layers of Features from Tiny Images
- Approximation by superpositions of a sigmoidal function
- Automatic differentiation in PyTorch
- Approximation capabilities of multilayer feedforward networks
- Towards Evaluating the Robustness of Neural Networks
- Deep Learning Face Attributes in the Wild
- The sample complexity of pattern classification with neural networks: the size of the weights is more important than the size of the network
- Unitary Triangularization of a Nonsymmetric Matrix