引導式測驗 · 中等 等級
人工智慧決策測驗的反事實解釋
區分基於模型的反事實和因果保證,並評估可行性、可操作性、約束和訴求需求。
問題 1 的 8
What question does a counterfactual explanation answer?
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 question does a counterfactual explanation answer?
- What small input change would make the model produce a different output?
- What source code created the model?
- What decision will a court make?
- What data are stored on a server?
- Does a counterfactual prove that changing a feature will cause a real-world outcome?
- 是的,總是
- No; it describes model behavior, not necessarily real-world causation
- Only for credit decisions
- Only if the model is interpretable
- A counterfactual says a person would qualify if they were ten years younger. What concern does that raise?
- It is necessarily a helpful action
- Age is always easy to change
- The change may be immutable, infeasible, or legally problematic
- It reveals the full model source code
- Why constrain counterfactual feature combinations?
- To avoid examples that are implausible or impossible
- To make all explanations identical
- To prevent model validation
- To increase the number of protected features
- How should a counterfactual about a loan balance be presented?
- As a guaranteed loan approval
- As proof of the lender’s causal reasoning
- As a model-based contrast that is not a promise of approval
- As a replacement for all adverse-action duties
- Which quality matters when a counterfactual is meant to help a person take action?
- Feasibility and actionability
- A large number of technical terms
- 模型的訓練時長
- Whether it hides errors
- What should happen if a counterfactual recommends changing a protected or immutable trait?
- Treat it as ordinary advice
- Ignore it without review
- Escalate it as a possible fairness or legal design issue
- Tell the person the change is mandatory
- Which process complements counterfactual explanations in consequential decisions?
- Data correction, human review, and appeal routes
- Removing all decision records
- Relying only on the generated explanation
- Making the model score confidential