知识卡片

普通读书笔记卡 · 8968.a · 生成模型:GPT-6

压缩要配合信息保留目标

理解等级:中级

只惩罚表示复杂度,最容易得到空向量或所有样本相同的退化解。因此压缩必须与重建、预测或其他信息保留约束配合。低维只是容量约束的一种;连续向量的维数不能直接等同有限比特率,实际存储还取决于精度和编码。

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

      Many representation learning algorithms capture these goals by penalizing or constraining the complexity of the learned representation. Such a penalty must be paired with some other objective or else it will lead to a degenerate solution: just make the representation contain nothing at all. Usually we try to compress the signal in certain ways while preserving other aspects of the information it carries. Autoencoders and

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第30章《表示学习》(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/c030b.xhtml) - 结论依据:摘录中关于压缩要配合信息保留目标的定义、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:Many representation learning algorithms capture these goals by penalizing or constraining the complexity of the learned representation. Such a penalty must be paired with some other objective or else it will lead to a degenerate solution: just make the representation contain nothing at all. Usually we try to compress the signal in certain ways while preserving other aspects of the information it carries. Autoencoders and