TL; DR
Image classification model (probably vision models in general) are spatially brittle. Small combination of rotation and translation can radically change model output.
Introduction
The focus in on spatial robustness, i.e. small spatial transfor...
