引導式測驗 · 中等 等級

What AI Accuracy Percentages Really Mean Quiz

Covers accuracy definitions, held-out evaluation, class imbalance, confusion-matrix metrics, subgroup performance, and deployment drift.

問題 1 的 8

A test set has 100 examples and a system gets 95 right. What does 95 percent accuracy describe?

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。

  1. A test set has 100 examples and a system gets 95 right. What does 95 percent accuracy describe?
    • Its measured fraction correct on that evaluated set
    • A guarantee for every future input
    • The percentage of positive predictions that are correct
    • The false-positive rate
  2. A dataset contains 99 negative cases and one positive. A model predicts negative for every case. What accuracy does it get?
    • 50 percent
    • 99 percent
    • 1 percent
    • 0 percent
  3. What does recall measure?
    • The share of predicted positives that are correct
    • The fraction of all cases that are negative
    • The share of actual positives the model finds
    • The model's latency
  4. Why keep a final test set separate from iterative tuning?
    • It makes the training set larger
    • It removes all class imbalance
    • It guarantees future accuracy
    • Repeated tuning on it can leak information and inflate apparent performance
  5. What can an aggregate accuracy score hide?
    • Different error rates across important subgroups
    • The number of cases tested
    • Whether any prediction was made
    • The model's file format
  6. Which evaluation detail is essential for interpreting a reported score?
    • The marketing slogan
    • The task, test data, date, and error measures
    • The color of the chart
    • The model's nickname
  7. Why can old benchmark performance become less relevant after deployment?
    • Accuracy is always zero after launch
    • The test set changes the model automatically
    • The real data or conditions may shift
    • Subgroups no longer exist
  8. What should happen when a high-consequence model output is uncertain or wrong?
    • Treat the percentage as proof
    • Hide the output from all auditors
    • Ignore the affected person
    • Use a defined review and appeal path