Reproduce “The False Promise of Zero-Shot Super-Resolution in Machine-Learned Operators” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mansi Sakarvadia, Kareem Hegazy, Amin Totounferoush, Kyle Chard, Yaoqing Yang, Ian Foster, Michael W. Mahoney
An empirical study investigates the claimed zero-shot super-resolution capabilities of Machine-Learned Operators (MLOs) like the Fourier Neural Operator, finding that these models often fail due to aliasing when generalizing to resolutions outside their training distribution. The work proposes and validates a multi-resolution training protocol, demonstrating it effectively enables robust inference across varied data resolutions.
Key insights
- MLOs' architectural design alone does not guarantee robust zero-shot multi-resolution inference, consistently failing due to aliasing when generalizing to different resolutions.
- Neither physics-informed optimization constraints nor existing band-limited learning approaches effectively resolve the zero-shot multi-resolution generalization problem.
- Robust multi-resolution performance for MLOs is best achieved by training on diverse resolutions, potentially using datasets with a majority of computationally cheaper low-resolution data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.