The continued growing interest in subseasonal prediction across scientific sectors (e.g., government operations, research, and private industry) has led to the creation of large and high quality datasets over the last decade to assess predictability on time scales between weather and climate, and is a testament to the societal need for a deeper understanding of predictive capabilities at such time scales. Despite recent progress, subseasonal prediction remains particularly challenging because the sources of predictability are limited; predictability stemming from atmospheric initial conditions is substantially reduced beyond approximately two weeks and the ocean generally does not offer added predictability until a trajectory reaches the seasonal timescale. The ability of machine learning to capture complex, non-linear patterns in the climate system that are difficult for humans to identify and understand presents a major opportunity to advance our predictive capabilities at subseasonal timescales. Here we present two studies: one focusing on the use of unsupervised learning to uncover large-scale subseasonal weather regimes and assess their predictability, and the second study focuses on the use of deep learning for bias correction of global subseasonal temperature and precipitation forecasts. Both of these studies use an ensemble set of reforecasts generated from a coupled, initialized subseasonal prediction system. These studies demonstrate how machine learning can be used to extend Earth system prediction and uncover hidden signals in initialized reforecasts, including how ensemble members can disperse to then largely agree again on a subseasonal prediction at a later lead time, which we term "ensemble realignment."