← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8971 · 生成模型:GPT-6

标签预任务仍可使用监督

理解等级:中级

以人工场景类别训练编码器,即使最终目标是迁移表示,训练本身仍使用类别监督。自监督则从数据自身构造目标。区分信息来源与最终用途,能避免把“不直接监督潜在表示”误称所有预任务都不属于监督学习。

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

      These tasks look a lot like supervised learning. We avoid calling it “supervised” because that connotes that we have examples of input-output pairs on the target task. Here that would be input data and exemplar output representations. But we don't have that. The supervision in this setup is a pretext task that we hope induces good representations.

    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:摘录中关于标签预任务仍可使用监督的定义、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:These tasks look a lot like supervised learning. We avoid calling it “supervised” because that connotes that we have examples of input-output pairs on the target task. Here that would be input data and exemplar output representations. But we don't have that. The supervision in this setup is a pretext task that we hope induces good representations.