← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8928 · 生成模型:GPT-6

学习不能保证消除别名

理解等级:中级

多通道学习可能找到滤波器来缓解抽取的混叠,甚至利用其他通道补偿,但损失函数未必惩罚所有别名伪影。依靠训练自动解决会占用表示能力且没有保证。分析步长层应检查频响与移位稳定性,而不只看训练准确率。

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

      One can argue that these artifacts might not be important when the kernel is being learned. Indeed, the learning could search for kernels that minimize the artifacts due to aliasing as those probably increase the loss. Also, as each layer is composed of many channels, the set of learned kernels could learn to compensate for the aliasing produced by other channels. However, this reduces the

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第24章《卷积神经网络》(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/c024b.xhtml) - 结论依据:摘录中关于学习不能保证消除别名的定义、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:One can argue that these artifacts might not be important when the kernel is being learned. Indeed, the learning could search for kernels that minimize the artifacts due to aliasing as those probably increase the loss. Also, as each layer is composed of many channels, the set of learned kernels could learn to compensate for the aliasing produced by other channels. However, this reduces the