Title:  Self-supervised Latent Space Factorization for Image Segmentation with Application to Seismic Interpretation

Committee: 

Dr. AlRegib, Advisor     

Dr. Frakes, Chair

Dr. Hoffman

Abstract: The objective of the proposed research is to explore the latent space of deep learning models with the goal of associating latent variables with semantic information in images. In most deep learning frameworks, focus is placed on end-to-end training of models, usually with the goal of beating benchmarks. Hence, the feature space is packed with more convolutional layers and various heuristics to achieve better benchmarks. This makes latent variables of deep learning frameworks highly redundant, resulting in low per-parameter accuracy. In our preliminary research, we explored how to separate semantic representations in images and associate them with latent variables. Using self-supervised frameworks, we show that we can control the latent space to generate meaningful reconstruction and segmentation of input images. We control the latent space by projecting variables to orthogonal latent subspaces and directing the projected variables to represent areas of interest in reconstructed images, resulting in weak self-supervised segmentation. In our proposed work, we show that our method could be generalized to learning and disentangling latent variables, for fine-grain control of semantic representation in generated images.