← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9006.a · 生成模型:GPT-6

增强不保证精确不变

理解等级:中级

训练中展示某类变换,可鼓励模型学会相应不变性或等变性,但有限样本、优化和模型能力并不保证所有参数取值下都严格满足它。与架构强制实现的对称性区别开,在变换强度及组合上分层评估,避免把训练策略当作形式保证。

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

      Simple transformations like mirror flipping or cropping can be very effective while also being easy to implement by hand. After learning, the model will be invariant (or sometimes equivariant, if the y values are transformed too) with respect to these transformations. But these are just a few of all possible transformations that we might want to be invariant to. Imagine if we are building a

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第36章《鲁棒性与通用性训练》(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/c036.xhtml) - 结论依据:摘录中关于增强不保证精确不变的定义、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:Simple transformations like mirror flipping or cropping can be very effective while also being easy to implement by hand. After learning, the model will be invariant (or sometimes equivariant, if the y values are transformed too) with respect to these transformations. But these are just a few of all possible transformations that we might want to be invariant to. Imagine if we are building a