引導式測驗 · 硬 等級
可解釋性與準確性權衡測驗
了解可解釋性和準確性何時會發生衝突、如何測試兩者,以及為什麼 NIST 將可信度視為依賴上下文的可信度。
問題 1 的 8
How does NIST describe possible tradeoffs between interpretability and predictive accuracy?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- How does NIST describe possible tradeoffs between interpretability and predictive accuracy?
- They may arise in some contexts, but are not inevitable
- Interpretability always reduces accuracy
- Accuracy proves interpretability
- The two terms mean the same thing
- Why should a team measure performance by relevant groups as well as overall?
- Group metrics always increase accuracy
- NIST requires identical results for every person
- An aggregate can hide different error rates or failures across populations
- It replaces testing on deployment data
- What does an explanation from a post-hoc tool establish by itself?
- It proves the prediction is correct
- It offers a model-dependent account that needs validation, not proof of causality or fairness
- It guarantees equal outcomes
- It reveals every fact about the training data
- Which set of qualities does NIST treat as distinct trustworthiness characteristics?
- Accuracy and interpretability only
- Transparency as a substitute for fairness
- One universal trust score
- Validity/reliability, explainability/interpretability, privacy, and fairness among others
- What should define an acceptable balance between accuracy and explanation quality?
- A vendor slogan
- The use context, risks, affected people, and measured performance
- The model’s parameter count alone
- An assumption that the simplest model is always best
- Which statement describes the status of the NIST AI RMF?
- A US federal statute imposing one required metric
- An EU regulation
- Voluntary risk-management guidance
- A model certification label
- What can a transparent model still do?
- Be inaccurate, biased, or poorly validated
- Guarantee fair results
- Eliminate uncertainty
- Replace human judgment in every context
- When should a team revisit the tradeoff assessment?
- Only before the first training run
- Never, once accuracy is measured
- Only when a vendor changes its logo
- After updates, shifts in the population, or changes in use context