Bayesian Scattering: A Principled Baseline for Uncertainty on Image Data
This provides a principled baseline for uncertainty quantification in image data, addressing a gap in the field for interpretable methods, though it is incremental as it adapts existing scattering transforms to a Bayesian framework.
The authors tackled the lack of interpretable, mathematically grounded baselines for uncertainty quantification in image data by proposing Bayesian scattering, which combines a wavelet scattering transform with a probabilistic head to provide sensible uncertainty estimates even under distribution shifts, achieving solid baseline performance across medical imaging, wealth mapping, and molecular property optimization tasks.
Uncertainty quantification for image data is dominated by complex deep learning methods, yet the field lacks an interpretable, mathematically grounded baseline. We propose Bayesian scattering to fill this gap, serving as a first-step baseline akin to the role of Bayesian linear regression for tabular data. Our method couples the wavelet scattering transform-a deep, non-learned feature extractor-with a simple probabilistic head. Because scattering features are derived from geometric principles rather than learned, they avoid overfitting the training distribution. This helps provide sensible uncertainty estimates even under significant distribution shifts. We validate this on diverse tasks, including medical imaging under institution shift, wealth mapping under country-to-country shift, and Bayesian optimization of molecular properties. Our results suggest that Bayesian scattering is a solid baseline for complex uncertainty quantification methods.