知识卡片

普通读书笔记卡 · 9010.b · 生成模型:GPT-6

蒸馏学习软类别分布

理解等级:高级

对 K 类分类,教师输出 K 个非负且和为一的概率,落在 K−1 维单纯形 Δ^(K−1),学生最小化与它的差异。目标交叉熵 H(t(x),f(x)) 在固定教师下等价于最小化 KL(t||f)。K−1 是自由度,不是概率向量只有 K−1 个元素。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      Knowledge distillation is a simple and popular algorithm for learning a classifier from a teacher. We assume that the teacher is a well-trained classifier that outputs a K – 1-dimensional probability mass function (pmf), t : X →△K–1 (see section 9.7.3). The student, a model fθ : X →△K–1, simply tries to imitate the teacher's output pmf: for each input x that the student sees,

    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:摘录中关于蒸馏学习软类别分布的定义、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:Knowledge distillation is a simple and popular algorithm for learning a classifier from a teacher. We assume that the teacher is a well-trained classifier that outputs a K – 1-dimensional probability mass function (pmf), t : X →△K–1 (see section 9.7.3). The student, a model fθ : X →△K–1, simply tries to imitate the teacher's output pmf: for each input x that the student sees,