Fundamentals of testing AI-based systems · Question 42 of 60

A team tests a locked image classification model on GPUs and in rare cases receives slightly different outputs for identical inputs. How is this to be classified?

1 answer is correct. Hard

1. A The model must be adaptive, because locked models always return the same for the same input
2. B The test environment is defective, because locked systems deliver exactly the same in every environment and on all hardware
3. C Even locked systems with large networks can be non-deterministic due to floating-point limits and GPU parallelism
4. D The model keeps learning unnoticed during the test and changes its weights