Signal Recovery from Pooling Representations

Abstract

In this work we compute lower Lipschitz bounds of ${\ell }_p$ pooling operators for $p=1,2,\infty$ as well as ${\ell }_p$ pooling operators preceded by half-rectification layers. These give sufficient conditions for the design of invertible neural network layers. Numerical experiments on MNIST and image patches confirm that pooling layers can be inverted with phase recovery algorithms. Moreover, the regularity of the inverse pooling, controlled by the lower Lipschitz constant, is empirically verified with a nearest neighbor regression.

17 pages, 3 figures

Cited by in corpus (13)

• Optimizing the Latent Space of Generative Networks
• Understanding and Improving Convolutional Neural Networks via Concatenated Rectified Linear Units
• i-RevNet: Deep Invertible Networks
• Efficiently utilizing complex-valued PolSAR image data via a multi-task deep learning framework
• Unsupervised Feature Learning from Temporal Data
• The Sparse Manifold Transform
• Unsupervised Learning of Spatiotemporally Coherent Metrics
• Globally Injective ReLU Networks
• Analysis of Invariance and Robustness via Invertibility of ReLU-Networks
• Towards Learning a Self-inverse Network for Bidirectional Image-to-image Translation
• TunaGAN: Interpretable GAN for Smart Editing
• Learning Boolean functions with concentrated spectra
• Data-driven Regularization via Racecar Training for Generalizing Neural Networks