ISTQB CT-AI
ML Model Testing: Metamorphic, Drift, A/B
Metamorphic testing
- What is metamorphic testing (MT)? Easy
- What does a metamorphic relation (MR) describe? Easy
- When is metamorphic testing particularly useful? Easy
- In which AI application areas has metamorphic testing already been used? Easy
- Which test objective of metamorphic testing describes that outputs remain stable under perturbations of the input? Easy
- From what do testers derive metamorphic relations? Easy
Drift and model fit
- What is data drift? Easy
- What is concept drift? Easy
- Which is an example of data drift? Easy
- What does dynamic drift testing rely on? Easy
- What is overfitting? Easy
- What is underfitting? Easy
- When should testing for overfitting and underfitting take place? Easy
- What does static drift testing compare? Easy
Comparative testing
- What is A/B testing? Easy
- Which is a simple example of A/B testing? Easy
- Which statistical techniques are most popular for A/B testing? Easy
- What is back-to-back testing? Easy
- What needs to be generated for back-to-back testing? Easy
- Why should the pseudo-oracle and the system under test ideally not share common software components? Easy
Metamorphic testing
- When is metamorphic testing typically selected over traditional oracle-based testing? Medium
- How can metamorphic relations be validated? Medium
- Which limitations does metamorphic testing have? Medium
- Which is an example of a monotonic metamorphic relation? Medium
- For which types of testing can metamorphic testing be used? Medium
- What is checked in metamorphic testing when a function is too complex to generate expected results? Medium
Drift and model fit
- Which is an example of concept drift? Medium
- In a film recommendation system, what is direct and what is indirect user feedback? Medium
- How is it determined in dynamic drift testing whether drift is present? Medium
- How is overfitting tested for? Medium
- How can overfitting and underfitting be detected during testing? Medium
- How does underfitting show in the learning curves? Medium
- Which values indicate underfitting? Medium
Comparative testing
- How does A/B testing help with the test oracle problem? Medium
- Which precondition is needed to test updates of an AI system with A/B testing? Medium
- How can A/B testing be used for self-learning systems? Medium
- Which statements about the pseudo-oracle in functional back-to-back testing are correct? Medium
- In which situations does back-to-back testing provide particular value? Medium
- What is a significant difference between A/B testing and back-to-back testing? Medium
- Where can the test inputs for back-to-back testing come from? Medium
Metamorphic testing
- An image classifier has correctly identified a photo as "cat". Which follow-up test cases result from an invariance relation? Hard
- A creditworthiness model is to be tested. Which metamorphic relation is a monotonicity relation? Hard
- A search engine returns 800 hits for "running shoes". Which follow-up test case results from a sensible metamorphic relation derived from domain knowledge? Hard
- A team derives metamorphic follow-up test cases from a source test case that has itself failed. What is the problem? Hard
- For an insurance model, a team uses the relation "higher age always results in a higher premium" and finds no defects. Later it turns out that age and pre-existing conditions interact. What happened here? Hard
- A team plans to test a route optimization system exclusively with metamorphic testing. What is recommended? Hard
- A team derives several metamorphic relations for an ML system. Which statements about this are correct? Hard
- A speech recognition system has correctly transcribed a sentence. Which follow-up test cases can be derived from metamorphic relations whose expected result is the same transcript? Hard
Drift and model fit
- Which mappings of situation and type of drift are correct? Hard
- A fraud detection system receives no feedback at all on whether its decisions were correct. How can drift still be tested? Hard
- A streaming service wants to check whether its recommendation model is drifting and uses the users' viewing data for this. Which statements are correct? Hard
- A model achieves 95% accuracy on the validation dataset but only 71% on a separate test dataset with less common examples. What can be concluded from this? Hard
- A model achieves 61% accuracy on the training data and 59% on the validation data; even with more epochs, the values hardly improve. Which statements are correct? Hard
Comparative testing
- A city wants to check whether an update to its AI-based traffic routing system improves journeys. Which approach fits? Hard
- A team expects an A/B test to provide the test cases for a new model. What is there to say about this? Hard
- A self-learning pricing system makes a change overnight. What does the use of A/B testing look like here? Hard
- As a pseudo-oracle for back-to-back testing, a team builds a second model with the same open-source library and the same algorithm. Both models always deliver matching results. What must be considered? Hard
- A team is to build a pseudo-oracle for functional back-to-back testing on a tight budget. Which statements help? Hard
- An ML system is migrated from the development to the production environment. The team wants to ensure that it behaves the same there, especially for unusual inputs. Which technique is suitable? Hard
- A team wants, on the one hand, to find out which of two model variants has the better accuracy and, on the other, to detect defects in a new implementation. Which techniques fit each purpose? Hard