Researchers training image-recognition systems found that systems achieving high accuracy on standard test sets sometimes performed poorly on images differing only slightly — for instance, photographs of the same objects taken under unusual lighting or from non-standard angles. The same systems also occasionally classified deliberately altered images, imperceptibly modified to humans, into entirely incorrect categories with high confidence.
Which inference is most strongly supported by the passage?
- A
High test-set accuracy reliably indicates robust performance across all related conditions
- B
Image-recognition systems are unable to perform any classification correctly
- Ccheck_circle
Apparent competence on benchmark tasks may not generalize to inputs outside the training distribution
- D
Human observers cannot perceive any differences between altered and original images
Explanation
Failure on slight variations and adversarial alterations supports B. A is contradicted; C overstates; D contradicts the imperceptibility note about humans, not the passage's broader claim.