知识卡片
软标签包含类别相似信息
相比 one-hot 标签,教师分布可在其他类别上分配概率,使学生接收到样本与相近类别的关系,例如家猫也有部分虎类概率。额外信号在某些条件下可加快学习,同时也会携带教师错误和偏差;不能把更细腻的概率解释为更接近真实分布的必然保证。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 迁移学习与适配
原书出处 · 1 处
- · 查看原书依据
Notice that this learning problem is very similar to softmax regression on labeled data, and a diagram comparing the two is given in figure 37.4. The difference is that in knowledge distillation the targets are not one-hot (like in standard label prediction problems) but rather are more graded. The teacher outputs a probability vector that reflect the teacher's belief as to the class of the
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第37章《迁移学习与适配》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c037.xhtml)
- 结论依据:摘录中关于软标签包含类别相似信息的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:Notice that this learning problem is very similar to softmax regression on labeled data, and a diagram comparing the two is given in figure 37.4. The difference is that in knowledge distillation the targets are not one-hot (like in standard label prediction problems) but rather are more graded. The teacher outputs a probability vector that reflect the teacher's belief as to the class of the