We present Stacked U-Nets, and neural net architecture that combines the information globalization properties of multigrid solvers with the expressive power of neural nets. Stacked U-Nets achieve state of the art performance on natural image segmentation (the Pascal VOC benchmark) using an order of magnitude fewer parameters than more complex models.
This presentation is part of Minisymposium “MS36 - Computational Methods for Large-Scale Machine Learning in Imaging (2 parts)”
organized by: Matthias Chung (Virginia Tech) , Lars Ruthotto (Department of Mathematics and Computer Science, Emory University) .