引導式測驗 · 中等 等級
高風險決策測驗中的可解釋模型與黑盒子
將本質上可解釋的模型與事後解釋進行比較,並評估準確性、保真度、任務適合度和高風險監督。
問題 1 的 8
What distinguishes an inherently interpretable model from a black box with a post-hoc explanation?
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。
- What distinguishes an inherently interpretable model from a black box with a post-hoc explanation?
- The interpretable model exposes its computation directly; the post-hoc method approximates a separate model
- The interpretable model never makes errors
- A black box always has no inputs
- Post-hoc explanations reveal source code
- What does Cynthia Rudin recommend for high-stakes decisions when performance is adequate?
- Use the most complex model available
- Prefer an inherently interpretable model over relying on a post-hoc explanation
- Avoid evaluating model accuracy
- Replace all human review
- Is “simple models always match black-box accuracy” a valid conclusion?
- Yes, for every task
- Only in medical tasks
- No; model choice depends on task and validation evidence
- Yes, if a rule list is used
- Why can a post-hoc explanation be risky?
- It may not faithfully represent the black-box model’s actual behavior
- It always reveals too much source code
- It prevents model testing
- It is identical to the model itself
- Does an interpretable model automatically guarantee fairness?
- Yes, transparency removes bias
- No; it can still use biased data or a discriminatory target
- Only if the model is linear
- Yes, if it is small
- How should candidate models be compared?
- On separate, unrelated test cases
- Only by vendor marketing
- On the same held-out cases using task-relevant metrics and error analysis
- By number of parameters alone
- What should a team test about a post-hoc explanation?
- Whether it tracks actual predictions when inputs change
- Whether it uses a colorful chart
- Whether it avoids every technical term
- Whether it is longer than the model card
- What supports using a black box in a high-stakes setting?
- A claim that it is state of the art
- Evidence it is necessary for the task, validation, limits on authority, monitoring, and meaningful review
- A feature-importance chart alone
- A statement that no errors occurred