← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8973.c · 生成模型:GPT-6

正负配对定义表示差异

理解等级:中级

正对规定哪些变化应被视为相同,负对规定哪些样本要区分。随机取不同样本作为负对,可能把同语义实例当作假负例;变换过强也可能使正对不再共享目标语义。配对规则因此是模型假设,而不是中立的数据准备。

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

      Contrastive learning fixes this issue by coupling an alignment loss with a second loss that pushes apart embeddings of datapoints for which we do not want an invariant representation. The supervision for contrastive learning comes in the form of positive pairs and negative pairs. Positive pairs are two datapoints we wish to align in z-space; if we wish for invariance to T then a positive

    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:摘录中关于正负配对定义表示差异的定义、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:Contrastive learning fixes this issue by coupling an alignment loss with a second loss that pushes apart embeddings of datapoints for which we do not want an invariant representation. The supervision for contrastive learning comes in the form of positive pairs and negative pairs. Positive pairs are two datapoints we wish to align in z-space; if we wish for invariance to T then a positive