引導式測驗 · 中等 等級

資料集中的代表性與取樣偏誤測驗

確定資料集覆蓋範圍差距,了解子組評估,並解釋總體分數和經典臉部分析審核的限制。

問題 1 的 8

A training set contains few examples from a group the model will serve. Which concern is most direct?

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 training set contains few examples from a group the model will serve. Which concern is most direct?
    • Representation bias
    • Encryption bias
    • Deployment latency
    • 標籤洩漏
  2. Why can overall accuracy hide a representation problem?
    • The test set may be dominated by common examples
    • Overall accuracy always measures every subgroup equally
    • The model cannot calculate error rates
    • Group sizes never affect evaluation
  3. What did Gender Shades report about two facial-analysis benchmarks?
    • They contained only dark-skinned women
    • They were overwhelmingly composed of lighter-skinned subjects
    • They measured medical risk, not images
    • They were collected after 2025
  4. Which group had the highest misclassification rates in the paper’s evaluated systems?
    • Lighter-skinned men
    • All groups had identical rates
    • Darker-skinned women
    • People using wheelchairs
  5. Why should the Gender Shades result be described with scope limits?
    • It tested specific commercial systems and a binary gender-classification task at that time
    • It proved all AI products fail today
    • It measured every kind of face technology
    • It evaluated voice assistants
  6. Which factor can create a sampling problem after data collection begins?
    • Only collecting from people who complete a task or respond to a survey
    • A larger hard drive
    • A new user-interface color
    • A model-card title
  7. What should accompany subgroup performance estimates?
    • Only the overall score
    • Counts and uncertainty
    • The vendor’s logo
    • A claim of universal accuracy
  8. Do synthetic examples automatically solve authentic coverage gaps?
    • Yes, they replace all real users
    • Only for audio
    • No; synthetic examples do not automatically substitute for authentic coverage
    • Yes, if generated at scale