← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8984.a · 生成模型:GPT-6

梯度为零不证明分布相同

理解等级:中级

当模型与数据分布相同时,能量梯度的正负相在期望中抵消。但反向推断不成立:受限模型、局部驻点或梯度统计不充分,也可能得到零更新。有限采样导致的一次近零梯度,更不能证明已经完美拟合真实分布。

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

      The last line should make clear the intuition: we establish a contrast between data samples and current model samples, then update the model to decrease this contrast, bringing the model closer to the data. Once x(i) and are identically distributed, this gradient will be zero (in expectation); we have perfectly fit the data and no further updates should be taken.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第32章《生成模型》(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/c032.xhtml) - 结论依据:摘录中关于梯度为零不证明分布相同的定义、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:The last line should make clear the intuition: we establish a contrast between data samples and current model samples, then update the model to decrease this contrast, bringing the model closer to the data. Once x(i) and are identically distributed, this gradient will be zero (in expectation); we have perfectly fit the data and no further updates should be taken.