Is Zero-Shot Super-Resolution Possible in Operator Learning?
Researchers have theoretically investigated the phenomenon of zero-shot super-resolution in operator learning, where models trained on coarse grids can predict on finer grids without retraining. The study reveals that this capability can be information-theoretically impossible in certain benign scenarios, such as with rank-one linear operators. However, the research identifies H"older smoothness of output functions as a sufficient condition for successful zero-shot super-resolution and provides corresponding generalization bounds. AI
IMPACT Provides theoretical understanding for a key capability in operator learning models.