引导式测验 · 中等 等级

主动学习测验

检查标签循环、预言、不确定性、委员会分歧、多样性、专家评审和标签效率的限制。

相关引导路径主动学习
问题 1 的 8

一个团队拥有大量未标注的数据池，但专家时间有限。主动学习要求团队接下来标注什么？

预览每个问题8 问题

当您参加测验时，就会出现答案和解释。

  1. 一个团队拥有大量未标注的数据池，但专家时间有限。主动学习要求团队接下来标注什么？
    • Every example in a fixed order
    • Only examples the model already predicts with high confidence
    • Examples selected as informative for the current model
    • Only examples generated synthetically by the model
  2. 在指南的主动学习循环中，什么是预言机（oracle）？
    • The human expert or other source that supplies requested labels
    • The unlabeled pool before any selection
    • The model’s probability score for a class
    • The evaluation metric computed after training
  3. 两个产品分类器对某个类别存在严重分歧。哪种策略会优先考虑标注该示例？
    • Random data augmentation
    • 渐变裁剪
    • Label smoothing
    • Query-by-committee
  4. 分类器为某个未标注项分配了几乎相等的概率。为什么不确定性采样（uncertainty sampling）会查询它？
    • Its label has already been confirmed
    • The small probability margin signals uncertainty about its label
    • It is guaranteed to be mislabeled
    • It has the highest possible confidence
  5. 为什么批处理选择器（batch selector）会将多样性与不确定性结合起来？
    • To avoid repeatedly selecting uncertain cases from only one narrow region
    • To eliminate all human annotation
    • To make the model linear
    • To guarantee every selected label is correct
  6. A radiology team can afford only a small number of expert annotations. Which use best matches the guide?
    • Let the model diagnose without expert review
    • Label only scans already known to be easy
    • Prioritize ambiguous scans for radiologists to label, then assess the trained model
    • Claim that fewer labels always yield the same accuracy
  7. What happens after the oracle labels a selected batch in the pool-based loop?
    • The model discards its labeled set and starts over
    • The new labels are added to training and selection repeats
    • The unlabeled pool becomes a test set automatically
    • Human review is permanently removed
  8. Which statement best describes a successful active-learning result?
    • Every dataset always needs exactly half as many labels
    • The model chooses labels without any human involvement
    • Uncertainty sampling can never introduce bias
    • Comparable model performance may be reached with fewer paid annotations