Despite its status as a relatively new field, Computer Vision has experienced many fundamental shifts in the approaches and techniques it has explored -- e.g. from its early focus on signal processing to a more recent interest in optimization. However, as of 2021, few of these shifts seem as influential as the rise of Machine Learning in 2013. The combination of large amounts of labeled data, fast compute on GPUs, and learning algorithms' ability to process this data before the algorithm's actual application has led to substantial advancements for the field. Existing problems (like image recognition, segmentation, optical flow, etc.) have substantially better solutions and new, more ambitious problems (like learned scene rendering from video) have emerged. Reflecting this trend, this thesis will present several early works which integrate machine learning into existing algorithms, resulting in substantially improved performance, and one later work which completely replaces known algorithms with a learning technique, resulting in better performance for existing formulations of the problem and promising results for a previously difficult formulation. This thesis will present a brief summary of the beginnings of deep learning research and then dive into the three techniques presented, with a brief literature survey on each topic, a summary and explanation of the algorithm, comparisons against other works, and a brief discussion of the ethical implications (organized by environmental and societal) these techniques may have. The three topics will consist of: (1) a technique that uses learning concepts to improve the efficacy of MRF preprocessing algorithms, (2) a new algorithm for improving panorama creation, a.k.a. image stitching which utilizes machine learning modules, and (3) a learning approach for the task of Autofocus.