← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8971.b · 生成模型:GPT-6

自监督从数据构造目标

理解等级:中级

令输入为V1(X)、目标为V2(X),可从同一原始数据产生训练对,例如用过去帧预测未来、可见区域预测遮罩区域。标签生成规则依然携带人为选择;必须阻断目标泄漏,并检验模型是否只利用无关采样或处理痕迹。

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

      Predictive learning is great when we have good prediction targets that induce good representations. What if we don't have labeled targets provided to us? Instead we could try to cook up targets out of the raw data itself; for example, we could decide that the top right pixel's color will be the “label” of the image. This idea is called self-supervision. It looks like this:

    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:摘录中关于自监督从数据构造目标的定义、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:Predictive learning is great when we have good prediction targets that induce good representations. What if we don't have labeled targets provided to us? Instead we could try to cook up targets out of the raw data itself; for example, we could decide that the top right pixel's color will be the “label” of the image. This idea is called self-supervision. It looks like this: