Ukryty dług techniczny w quizie dotyczącym systemów ML

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. Który dalszy podmiot zależny jest konsumentem niezadeklarowanym?
   • A downstream component that uses a feature or prediction without tracked ownership
   • A model trained without labels
   • A missing value in a tensor
   • A GPU that lacks a driver
2. W jaki sposób model może stworzyć pętlę informacji zwrotnej w rankingu?
   • Its loss function changes the programming language
   • Its selections shape which future interactions become training data
   • Its predictions remove all user behavior
   • Its features become immutable
3. Co opisuje splątanie w systemie ML?
   • Training data contain only one class
   • All model parameters are independent
   • Changes in one component can have nonlocal effects through dependencies
   • The model is stored in a container
4. Dlaczego zmiana schematu funkcji może być ryzykowna, gdy kilka modeli korzysta ze wspólnego potoku?
   • It removes the need for integration tests
   • It automatically improves every model
   • Shared features cannot be versioned
   • The change can alter behavior for multiple consumers
5. Co umowy dotyczące schematów i własność mogą pomóc ujawnić?
   • The model's exact future accuracy
   • Changes in data meaning and responsibility across system boundaries
   • A guaranteed causal relationship
   • The optimal number of latent factors
6. Why may model-only monitoring miss an upstream data issue?
   • Monitoring disables prediction logs
   • The model cannot receive data
   • Input or pipeline failures can cause output changes before model code changes
   • Upstream systems are always stable
7. How can a team reduce risk when retiring an unused model?
   • Remove its evaluation report first
   • Delete the artifact without checking dependencies
   • Keep every model forever
   • Map consumers and use a staged retirement plan
8. Which symptom can indicate accumulating technical debt?
   • Changes become slower and rollback or regression analysis becomes difficult
   • Every model has fewer parameters
   • Training loss decreases
   • The repository contains tests