持续学习和灾难性遗忘测验

当您参加测验时，就会出现答案和解释。

1. 什么是“灾难性遗忘”？
   • 当学习一项新任务会大幅降低之前学习任务的表现时
   • 当模型内存不足并崩溃时
   • 当训练数据被意外删除时
   • 当模型拒绝做出预测时
2. 持续学习的核心权衡通常被称为：
   • 偏差-方差权衡
   • 稳定性-可塑性困境
   • 速度与精度的权衡
   • 探索-利用权衡
3. 弹性权重合并 (EWC) 有什么作用？
   • 删除不重要的神经元
   • 新任务的学习率加倍
   • 惩罚对之前学习的任务很重要的权重变化
   • 永久存储过去的每个训练示例
4. 基于重放的方法如何对抗遗忘？
   • 通过永久冻结整个网络
   • 通过使用更大的学习率
   • 通过删除损失函数
   • 通过将存储或生成的旧示例交错到新训练中
5. 梯度情景记忆（GEM）如何在学习新任务时限制梯度更新？
   • It projects the gradient so its first-order approximation does not increase loss on stored prior-task examples
   • It replaces all prior-task data with new examples
   • It freezes every parameter for all future tasks
   • It increases weights that were unimportant to earlier tasks
6. 为什么对新任务进行天真的微调会导致遗忘？
   • GPU 过热并损坏内存
   • 梯度下降将共享权重移向新任务，而没有保留旧任务的限制
   • 新任务总是有更多数据
   • 模型有意删除旧权重
7. How does iCaRL classify a test example after learning multiple classes?
   • By retrieving the single closest word in a text corpus
   • By selecting the class with the largest Fisher information
   • By comparing its feature vector with per-class means computed from stored exemplars
   • By replaying every training image at inference
8. Why is evaluating continual learning more than measuring final accuracy on the newest task?
   • New-task accuracy is the only relevant measure
   • Old tasks are removed from the evaluation
   • The model cannot be tested on multiple tasks
   • A method must also measure retention on prior tasks and its ability to learn new ones