Neural networks · Question 57 of 60
A test suite achieves 100% neuron coverage, yet the model fails in the field on slightly rotated images. Which explanations and consequences are there?
3 answers are correct. Hard
Show answer
Answer
- B Testers should additionally use techniques such as adversarial or metamorphic testing
- D Structural coverage does not guarantee good generalization or handling of real-world variations
- E The network may have learned spurious correlations and activate correctly for incorrect reasons
Explanation
Structural coverage alone does not guarantee that a network generalizes well or can handle real-world variations. Networks can learn spurious correlations and thus activate correctly for incorrect reasons. Testers should therefore use further techniques such as adversarial and metamorphic testing.