##Multiscale Hierarchical Convolutional Networks
Multiscale Hierarchical Convolutional Networks (hCNNs) are highly structured CNNs that formulate each layer as a multi-dimensional convolution. hCNNs provide a framework that allows to study and understand mathematical and semantic properties of deep convolutional networks.
Reference: J.-H. Jacobsen, E. Oyallon, S. Mallat, A.W.M. Smeulders; Multiscale Hierarchical Convolutional Networks.
Under submission, 2017
This repository is a preliminary release accompanying our arxiv submission. It will be updated to incorporate examples, compatibility with Tensorflow 1.0 / Keras 2.0 and more results soon.
Tested with: cudnn v5.1; cuda 8.0; Tensorflow 0.12; Keras 1.2.2
@article{jacobsen2017multiscale,
title={Multiscale Hierarchical Convolutional Networks},
author={Jacobsen, J{\"o}rn-Henrik and Oyallon, Edouard and Mallat, St{\'e}phane and Smeulders, Arnold WM},
journal={arXiv preprint arXiv:1703.04140},
year={2017}
}
- Understanding Deep Convolutional Networks, Mallat, 2016