知识卡片

普通读书笔记卡 · 9006.d · 生成模型:GPT-6

生成增强保标签需验证

理解等级:中级

算法 36.1 对每个 x_i 编码 z_i,重复 k 次采样 w_j∼N(0,σ²I),生成 g(z_i+w_j) 并沿用 y_i。实际实现应初始化并追加数据集,决定是否同时保留原样本。最重要的前提是重建与扰动不改变目标标签;算法复制标签本身不构成标签正确的证明。

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

      If we do this for a variety of settings of w, then we will get a set of output images that are all slight variations of one another, just like we saw previously in figure 33.15. One simple approach is to just use Gaussian perturbations in latent space, that is, w ∼ N (0, σ2) for some perturbation scale given by σ. Algorithm 36.1 formalizes

    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:摘录中关于生成增强保标签需验证的定义、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:If we do this for a variety of settings of w, then we will get a set of output images that are all slight variations of one another, just like we saw previously in figure 33.15. One simple approach is to just use Gaussian perturbations in latent space, that is, w ∼ N (0, σ2) for some perturbation scale given by σ. Algorithm 36.1 formalizes