Paired Autoencoders for Inference and Regularization
Emma Hart, Emory University
In this work, we describe a reduced approach that exploits technologies from machine learning (e.g., neural networks and auto-encoder networks) and dimensionality reduction models (e.g., low-rank and latent representations) to advance various technologies for inverse problems. We consider a decoupled approach for surrogate modeling, where unsupervised learning approaches are used to efficiently represent the input and target spaces separately, and a supervised learning approach is used to represent the mapping from one latent space to another. We demonstrate that our approach can outperform others in scenarios where training data for unsupervised learning is easily available, but the number of input/target pairs for supervised learning is small, and we introduce how the approach can be used for defining regularization or prior knowledge, and/or as a surrogate model for inversion, forward propagation, and adjoint free methods.