Appearance
Topic Drill 4.1.1: Locked and Adaptive AI-Based Systems
About this drill
LO AI-4.1.1 (K2): Compare the testability of locked and adaptive AI-based systems.
Lesson: Ch 4, section 4.1.1
This drill covers every point in section 4.1.1:
- Locked systems (fixed weights, retraining, lane detection / traffic sign examples) and adaptive systems (reward function, new environment, recommendations)
- GenAI chatbots between the two; the continuum
- Testability: locked = largely deterministic; an update = a new system; floating-point/GPU non-determinism
- Testing adaptive systems: before deployment (simulation, learning mechanism, controlled adaptation), automated core suite on significant changes, degradation thresholds and monitoring
Topic Drill: 4.1.1
Q1. A claims-triage model has frozen weights in production. Each quarter it is retrained on new data and the updated model is released. What does the syllabus say about testing each quarterly update?
AOnly the changed weights need to be tested
BNo testing is needed, because the system is locked
CThe updated locked system is typically considered a new system, and a fresh round of testing is required
DIt must be tested as an adaptive system, with continuous monitoring only
Q2. Where does the syllabus place LLM-based chatbots that are deployed as locked models at runtime but updated periodically?
ABetween fully locked and fully adaptive systems
BFully adaptive, because they generate new content
CFully locked, because the weights never change
DOutside the locked/adaptive categorization entirely
Q3. A locked image-classification DNN occasionally gives slightly different outputs for the same input when it runs on different GPU servers. How does the syllabus explain this?
AThe model is secretly self-learning
BThe test data is defective
CLocked systems are always perfectly deterministic, so this must be a hardware defect
DLarge neural networks can show non-deterministic behavior because of floating-point precision limits and hardware execution variations, especially with parallel computation or GPUs
Q4. Which of the following does the syllabus list as a way to test an adaptive AI-based system BEFORE deployment?
AFreezing its learning permanently
BSimulating changes in environmental conditions and testing its ability to adapt appropriately in controlled scenarios
CTesting only its user interface
DWaiting for production incidents
Q5. Why is testing an adaptive AI-based system more complex than testing a locked one?
ABecause adaptive systems have no requirements
BBecause adaptive systems cannot be automated
CBecause the system may change its behavior during the testing or as a result of the testing
DBecause adaptive systems are always larger
Q6. A deployed home-energy optimizer keeps adapting to each household. New behaviors cannot be predicted in advance. What does the syllabus recommend?
ABuild an automated test suite focused on the system's core functionality and run it whenever the system undergoes significant changes, to verify that adaptations are safe
BWrite test cases for every possible future behavior
CStop testing after deployment
DTest only the training data
Q7. Which TWO of the following are ADAPTIVE AI-based systems?
Select 2 options.
AA tax calculator built from predefined expert rules
BA smart-home system that keeps learning the occupants' routines and adjusts the heating
CAn OCR model whose weights were fixed at release
DA reinforcement learning warehouse robot that keeps optimizing its routes in production
EAn image classifier retrained once a year and then frozen
Q8. Besides testing adaptations for safety, what else does the syllabus say testing can check for an adaptive system in operation?
AThat the system has stopped learning
BThat the training data has been deleted
CThat every output exactly matches the original expected results
DThat performance has not degraded beyond a certain threshold, in response to significant changes or as part of ongoing monitoring
Q9. How does the syllabus describe the range of AI-based systems "in practice"?
AEvery system is either fully locked or fully adaptive
BThey span a continuum from fully deterministic, locked-down systems to deliberately non-deterministic, self-learning systems
CAll GenAI systems are fully adaptive
DAll safety-related systems are adaptive
0 / 9 answered
← 3.4.3 Neural Network Coverage · Next drill: 4.1.2 Statistical Approach →