The ability to identify distributional shifts in temporal data is a common theme for both change-point detection and various machine learning tasks. This thesis explores the interplay between machine learning and change-point detection, and uses this interplay to devise new methods that utilize change-point detection for improving machine learning and in the reverse case, utilizes machine-learning to devise improved change point detection methods. For improving change-point detection, we explore how supervision can help learn distance metrics for designing robust change-point detection methods. For improving machine learning, we explore how change-points can provide weak supervision for supervised machine learning tasks. We also investigate how machine learning models can adapt to distributional changes by learning to selectively mask time series channels with significant shifts.