课程学习测验

当您参加测验时，就会出现答案和解释。

1. What ordering does a common curriculum-learning strategy use?
   • Present examples from easier to harder according to a chosen schedule
   • Always present the most difficult example first
   • Sort examples alphabetically and never change the order
   • Use only examples that already have perfect predictions
2. 谁在 2009 年的一篇著名论文中引入了“课程学习”一词？
   • 杰弗里·辛顿
   • 约书亚·本吉奥及其同事
   • 吴恩达
   • 严乐存
3. 在自定进度学习中，谁来决定哪些示例足够“简单”，可以尽早进行训练？
   • 人类注释者手动标记难度
   • 硬件调度器
   • 模型本身，基于该模型具有低损失
   • 仅单独的预训练难度分类器
4. 为什么课程学习有助于优化？
   • 它自动提高学习率
   • 保证零训练误差
   • 它消除了对损失函数的需要
   • 它平滑了早期目标，有助于避免糟糕的局部最小值
5. How should a team test whether an easy-to-hard curriculum helps a model?
   • Compare against a suitable baseline with the same data and compute budget
   • Assume it helps whenever the schedule has more stages
   • Measure training loss once and skip held-out evaluation
   • Compare runs that use different data and different compute
6. 什么是“反课程”学习？
   • 完全没有示例的训练
   • 先展示更难的例子，然后再展示更简单的例子
   • 仅使用验证数据
   • 禁用反向传播
7. Which four tasks did the 2010 self-paced learning paper report in its experiments?
   • Speech translation, protein folding, weather prediction, and music synthesis
   • Image captioning, game playing, robotics, and medical diagnosis
   • Object localization, noun-phrase coreference, motif finding, and handwritten-digit recognition
   • Clustering, sentiment analysis, machine translation, and speech diarization
8. How can a curriculum be defined reproducibly?
   • Use the word “easy” without a criterion
   • Change the test set after each epoch
   • Select only examples the model already predicts correctly
   • Specify the example difficulty measure, ordering or schedule, and comparison baseline