SEANA module prevents noise shortcuts in self-supervised learning
November 20, 2026
Self-Aligned Noise Augmentation (SEANA) is a drop-in module that prevents encoders from using signal-dependent noise as a shortcut in self-supervised learning. It learns an invertible variance-stabilizing transform to resample independent noise, working without clean targets or changes to the underlying SSL objective.
HOW THIS AFFECTS YOU
●
researcherYou can use this to make self-supervised training more robust on noisy scientific imaging data like microscopy.