知识卡片
不同网络层保留不同相似性
书中彩色形状实验显示,较深编码对形状近邻更有用,而原始像素对颜色分类更有利。它说明层选择取决于任务,不能推出所有深层必然丢颜色或形状总优于颜色。此处是作者特定数据、网络与评估下的结果。
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
Next we will probe a bit deeper, and ask, how effective are these embeddings at classifying key properties of the data? On the right we show the accuracy of a one-nearest-neighbor color classifier (between the eight color classes) and a shape classifier (circle vs. triangle vs. square) applied on embedding vectors at each layer of the autoencoder's encoder. The zeroth layer corresponds to measuring distance
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:摘录中关于不同网络层保留不同相似性的定义、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:Next we will probe a bit deeper, and ask, how effective are these embeddings at classifying key properties of the data? On the right we show the accuracy of a one-nearest-neighbor color classifier (between the eight color classes) and a shape classifier (circle vs. triangle vs. square) applied on embedding vectors at each layer of the autoencoder's encoder. The zeroth layer corresponds to measuring distance