Lipschitz constant estimation of Neural Networks via sparse polynomial optimization

Abstract

We introduce LiPopt, a polynomial optimization framework for computing increasingly tighter upper bounds on the Lipschitz constant of neural networks. The underlying optimization problems boil down to either linear (LP) or semidefinite (SDP) programming. We show how to use the sparse connectivity of a network, to significantly reduce the complexity of computation. This is specially useful for convolutional as well as pruned neural networks. We conduct experiments on networks with random weights as well as networks trained on MNIST, showing that in the particular case of the ${\ell }_{\infty }$-Lipschitz constant, our approach yields superior estimates, compared to baselines available in the literature.

arXiv commentPublished as a conference paper in ICLR2020, originally submitted in September 25 2019 and available at https://openreview.net/forum?id=rJe4_xSFDB

References in corpus (7)

• Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)
• Provable defenses against adversarial examples via the convex outer adversarial polytope
• Safe Model-based Reinforcement Learning with Stability Guarantees
• Semidefinite relaxations for certifying robustness to adversarial examples
• Optimal Power Flow as a Polynomial Optimization Problem
• Wasserstein Adversarial Examples via Projected Sinkhorn Iterations
• Stability-certified reinforcement learning: A control-theoretic perspective

Cited by in corpus (15)

• Training robust neural networks using Lipschitz bounds
• Fixed Point Strategies in Data Science
• The Lipschitz Constant of Self-Attention
• Semialgebraic Optimization for Lipschitz Constants of ReLU Networks
• Communication-Efficient Robust Federated Learning Over Heterogeneous Datasets
• Certifying Incremental Quadratic Constraints for Neural Networks via Convex Optimization
• Markov-Lipschitz Deep Learning
• Semialgebraic Representation of Monotone Deep Equilibrium Models and Applications to Certification
• The Nonlinearity Coefficient - A Practical Guide to Neural Architecture Design
• Norm-based generalisation bounds for multi-class convolutional neural networks
• What training reveals about neural network complexity
• Analytical bounds on the local Lipschitz constants of affine-ReLU functions
• Building Compact and Robust Deep Neural Networks with Toeplitz Matrices
• Efficient Proximal Mapping of the 1-path-norm of Shallow Networks
• Depthwise Separable Convolutions Allow for Fast and Memory-Efficient Spectral Normalization