The authors partitioned ImageNet validation images using disagreements between a standard classifier and robustly trained classifiers, which are trained to resist input perturbations.
Brittle visual cues can mislead neural networks on unmodified images
Zhang and colleagues find that filtering or small geometric shifts can correct some image-classification errors while creating similar failures elsewhere.
Chinese Tech
Jiaming Zhang · Fuyao Zhang · Pengjun Xie · Wei Yang Bryan Lim
Nanyang Technological University · Alibaba Group
Research Digest··2 min read
Zhang and colleagues show that subtle visual patterns can hinder, rather than help, standard image classifiers on clean images.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.