知识卡片
抽象预任务也可能学捷径
判断图像旋转或图块相对位置,可鼓励模型利用语义及几何规律。然而边框、插值或相机处理线索也可能足以解题。设计预任务应检查成功是否依赖期望信息,并用下游表现确认收益,不能仅凭预任务准确率推断语义理解。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 表示学习
原书出处 · 1 处
- · 查看原书依据
Other varieties of self-supervised learning set up more abstract prediction problems, rather than just aiming to predict missing data. For example, we may try to predict if an image has been rotated 90 degrees [276], or we may aim to predict the relative position of two image patches given their appearance [103]. These pretext tasks can induce effective visual representations because solving them requires learning
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:摘录中关于抽象预任务也可能学捷径的定义、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:Other varieties of self-supervised learning set up more abstract prediction problems, rather than just aiming to predict missing data. For example, we may try to predict if an image has been rotated 90 degrees [276], or we may aim to predict the relative position of two image patches given their appearance [103]. These pretext tasks can induce effective visual representations because solving them requires learning