知识卡片

普通读书笔记卡 · 9006 · 生成模型:GPT-6

二维重叠不能证明同分布

理解等级:中级

图 36.2 用 CLIP 特征的前两个主成分展示,翻转、裁剪、颜色扰动使训练散点扩展并靠近另一来源。它是覆盖变化的直观示意;二维投影重叠不能证明完整图像分布相同,增强也不能凭空补齐训练中缺失的对象类别。应以目标任务评估检验收益。

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

      Figure 36.2: Data augmentation broadens the training distribution so that it might better cover the test cases. (a) Training data are from Caltech256 [175] and test data are from CIFAR100 [278]. The scatter plots show these data in a 2D feature space [the first two principle components (pcs) of CLIP [394]]. (b) Training data after mirror augmentations (random horizontal flips). (c) The same plus crops

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第36章《鲁棒性与通用性训练》(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/c036.xhtml) - 结论依据:摘录中关于二维重叠不能证明同分布的定义、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:Figure 36.2: Data augmentation broadens the training distribution so that it might better cover the test cases. (a) Training data are from Caltech256 [175] and test data are from CIFAR100 [278]. The scatter plots show these data in a 2D feature space [the first two principle components (pcs) of CLIP [394]]. (b) Training data after mirror augmentations (random horizontal flips). (c) The same plus crops