Hyperplane Arrangements of Trained ConvNets Are Biased
arXiv:2003.07797
Abstract
We investigate the geometric properties of the functions learned by trained ConvNets in the preactivation space of their convolutional layers, by performing an empirical study of hyperplane arrangements induced by a convolutional layer. We introduce statistics over the weights of a trained network to study local arrangements and relate them to the training dynamics. We observe that trained ConvNets show a significant statistical bias towards regular hyperplane configurations. Furthermore, we find that layers showing biased configurations are critical to validation performance for the architectures considered, trained on CIFAR10, CIFAR100 and ImageNet.
References in corpus (8)
- A Closer Look at Memorization in Deep Networks
- Predicting the Generalization Gap in Deep Networks with Margin Distributions
- Are All Layers Created Equal?
- An Empirical Study of Example Forgetting during Deep Neural Network Learning
- Deep ReLU Networks Have Surprisingly Few Activation Patterns
- Norm-Based Capacity Control in Neural Networks
- Complexity of Linear Regions in Deep Networks
- Geometry of Deep Convolutional Networks