Model Drift Detection Quiz

Answers and explanations appear as you take the quiz.

1. What is the difference between data drift and concept drift?
   • Data drift changes the inputs' distribution; concept drift changes the input-to-output relationship
   • They are two names for the same thing
   • Data drift only affects images; concept drift only affects text
   • Data drift happens in training; concept drift happens in testing
2. How should a team interpret a PSI value such as 0.30?
   • As a guarantee that model accuracy has fallen
   • As a signal to investigate under the chosen monitoring policy, not a universal proof of harmful drift
   • As an identical threshold for every feature and population
   • As proof that retraining must begin immediately
3. Why can input-based drift detection give an early warning?
   • Because inputs are always wrong first
   • Because inputs are smaller files
   • Because ground-truth labels often arrive much later than the predictions
   • Because models ignore labels entirely
4. What does a two-sample Kolmogorov-Smirnov test compare?
   • The average accuracy of two classifiers only
   • The class labels predicted by a single model
   • The parameter values of two neural networks
   • The empirical cumulative distribution functions of two samples
5. Which scenario is a clear example of concept drift?
   • What users consider 'spam' changes over time even for similar-looking emails
   • A camera produces slightly brighter images
   • Customer ages in the dataset stay the same
   • The dataset is stored in a new file format
6. What is a safe first response when an input-drift monitor triggers?
   • Automatically deploy a newly retrained model without evaluation
   • Investigate the data and impact, then decide whether retraining or another action is justified
   • Delete the reference data
   • Assume every shift is harmful concept drift
7. What can TensorFlow Data Validation check in an ML pipeline?
   • The causal effect of a feature
   • Whether every model prediction is correct
   • Data schemas and statistics, including unexpected changes
   • The absolute accuracy of an unlabeled dataset
8. What does the Population Stability Index (PSI) measure in a drift-monitoring setup?
   • The model’s accuracy on labeled examples
   • A causal effect of an input feature
   • A universal pass/fail threshold for all models
   • A numerical summary of how binned feature distributions differ between reference and current data