"ANNs can perform exceptionally well when new data resembles the training distribution, but their performance can deteriorate substantially under distribution shifts."
The WILDS research benchmark systematically evaluates machine-learning performance under realistic changes between training and deployment distributions. Across ten datasets covering applications including medical imaging, wildlife identification and satellite imagery, the researchers demonstrate significant performance degradation under distribution shift. This provides empirical support for the book's distinction between successful pattern recognition within familiar conditions and brittleness when the territory changes.
View Source