Self-Supervised Deep Learning (SSDL) is a challenging learning paradigm that seeks to unearth structure in an underlying data distribution without knowing ground truth labels or outputs for the intended downstream task. The simplicity of this paradigm's formulation makes it widely applicable to a host of problems wherein paired ground truth labels are intractable or expensive to acquire. Specifically in Computer Vision, SSDL brings a lot of value to problems wherein labels are either human-annotated, acquired by expensive sensors, or required at the per-pixel level. Recent advances in SSDL for Computer Vision have led to progress in several domains, such as semantic segmentation, classification, and other foundational vision tasks.

While there has been an impetus for SSDL in discriminative and regressive setups, Self-Supervised Deep Generative Models (DGMs) remain under-explored. Our endeavor in this work is to dive deep into SSDL and improve upon its utility for selected exciting applications. In particular, we tackle two problems. Namely, Self-Supervised Monocular Depth Estimation and Causal Deep Image Manipulation using a Self-Supervised Deep Generative Autoencoder. First, we show that incorporating long range per-pixel dependencies improves Self-Supervised Monocular Depth Estimation. Second, we infuse the notion of causality in a Self-Supervised Deep Generative Autoencoder to perform prior-free semantic segmentation of faces. Finally, we employ the developed semantic segmentation model to perform localized style editing in face images.