Predicting compressed representations of future data rather than raw values, enabling more robust and generalizable learning.