A. Cetin

Lower phase residual does not guarantee structural accuracy, but pairing untrained neural priors with explicit projection constraints cuts magnitude-free phase retrieval error by 29.3% without requiring any training data.

Achieving state-of-the-art stereo depth estimation with 2.65x faster inference and 2.46x lower compute costs using a novel token mixing approach.

A novel U-Net architecture slashes parameter count while boosting wildfire prediction accuracy by fusing Hadamard and Discrete Cosine Transforms in latent spaces.