知识卡片
重建好不保证任务距离好
即使自动编码器能重建输入,潜在空间的欧氏近邻仍需单独检验。重建损失并不唯一确定潜在坐标尺度和几何;编码解码可相互补偿。通过近邻可视化及下游分类观察表示是否保留任务相关相似性,比只看重建图片更有针对性。
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
We can test this in two ways: (1) for a given query, visualize the images in the dataset whose embeddings are nearest neighbors to the query's embedding, and (2) measure the accuracy of a one-nearest-neighbor classifier in embedding space. Below, in figure 30.4, we show both these analyses.
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:摘录中关于重建好不保证任务距离好的定义、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:We can test this in two ways: (1) for a given query, visualize the images in the dataset whose embeddings are nearest neighbors to the query's embedding, and (2) measure the accuracy of a one-nearest-neighbor classifier in embedding space. Below, in figure 30.4, we show both these analyses.