知识卡片

普通读书笔记卡 · 8975 · 生成模型:GPT-6

归一化限制向量尺度投机

理解等级:中级

单位长度嵌入使表示落在球面,点积与距离关系固定,避免简单通过无限放大向量拉开负对。零向量归一化需处理数值稳定。约束范数只规定尺度,不保证数据已经均匀铺开、语义可分或每个维度统计独立。

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

      Wang and Isola [494] showed that the contrastive loss (specifically the InfoNCE form) encourages two simple properties of the embeddings: alignment and uniformity. We have already seen that alignment is the property that two views in a positive pair will map to the same point in embedding point, that is, the mapping is invariant to the difference between the views. Uniformity comes from the negative

    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:摘录中关于归一化限制向量尺度投机的定义、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:Wang and Isola [494] showed that the contrastive loss (specifically the InfoNCE form) encourages two simple properties of the embeddings: alignment and uniformity. We have already seen that alignment is the property that two views in a positive pair will map to the same point in embedding point, that is, the mapping is invariant to the difference between the views. Uniformity comes from the negative