Multiresolution convolutional autoencoders

We propose a multi-resolution convolutional autoencoder (MrCAE) architecture that integrates and leverages three highly successful mathematical architectures: (i) multigrid methods, (ii) autoencoders (iii) transfer learning. The method provides an adaptive, hierarchical capitalizes on progressive training approach for multiscale spatio-temporal data. This framework allows inputs across multiple scales: starting from compact (small number of weights) network low-resolution data, our progressively deepens widens itself in principled manner to encode new information the higher resolution data based its current performance reconstruction. Basic learning techniques are applied ensure learned previous steps can be rapidly transferred larger network. As result, dynamically capture different scaled features at depths gains this adaptive illustrated through sequence numerical experiments synthetic examples real-world spatial-temporal