Fully hyperbolic convolutional neural networks

Convolutional neural networks (CNN) have recently seen tremendous success in various computer vision tasks. However, their application to problems with high dimensional input and output, such as high-resolution image video segmentation or 3D medical imaging, has been limited by factors. Primarily, the training stage, it is necessary store network activations for back-propagation. In these settings, memory requirements associated storing can exceed what feasible current hardware, especially 3D. Motivated propagation of signals over physical networks, that are governed hyperbolic Telegraph equation, this work we introduce a fully conservative high-dimensional output. We coarsening operation allows completely reversible CNNs using learnable discrete wavelet transform its inverse both coarsen interpolate state change number channels. show able achieve results comparable art 4D time-lapse hyper-spectral full segmentation, much lower footprint constant independent depth. also extend use variational auto-encoders, where optimization begins from an exact recovery discover level compression through optimization.