Table 1:

Accuracy of a ResNeXt-29, 8x64d trained using the full CIFAR-100 dataset (“Baseline”) and two synthetic experiments with altered datasets. (“Subsample”) drops 75% of the dolphin and mountain subclasses from the training dataset, and (“Random Noise”) assigns 25% of examples from these subclasses a random superclass label. Results reported are on superclass labels for the validation set. Numbers in parentheses are reductions in performance with respect to the baseline model for each experimental condition.

Subclass	Baseline Superclass	Baseline Subclass	Subsample Superclass	Subsample Subclass	Random Noise Superclass	Random Noise Subclass
Dolphin	0.69	0.78	0.65 (−4)	0.64 (−14)	0.67 (−2)	0.73 (−5)
Mountain	0.87	0.90	0.82 (−5)	0.71 (−19)	0.82 (−5)	0.73 (−17)