知识卡片
隐层物体选择性不是完整检测
书中场景分类和上色实例的部分通道对特定物体有选择性,说明完成预任务可能学到未被显式标注的结构。但高激活图不足以证明通道是完整、稳健的物体检测器,也不保证每个预任务都会出现同样表征或与人类机制一致。
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
The real power of these pretexts tasks lies not in solving the tasks themselves but in acquiring useful image embeddings z as a byproduct of solving the pretext task. The amazing thing that ends up happening is that z-space may have emergent structure that was not explicit in either the raw training data nor the pretext task. As a case study, consider the work of
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:摘录中关于隐层物体选择性不是完整检测的定义、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:The real power of these pretexts tasks lies not in solving the tasks themselves but in acquiring useful image embeddings z as a byproduct of solving the pretext task. The amazing thing that ends up happening is that z-space may have emergent structure that was not explicit in either the raw training data nor the pretext task. As a case study, consider the work of