← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8932.c · 生成模型:GPT-6

补充反例后仍需重新验证

理解等级:中级

发现失败样本后,可以将相关变化纳入训练并重训,再重复机制分析和域外测试。加入一种反例通常只修补一部分分布缺口,不能据此宣布对抗鲁棒性已解决。需明确扰动约束、标签语义与独立评估,避免把测试集改造成训练依据而仍当作未见测试。

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

      For complex architectures, adversarial examples are obtained as an optimization problem: What is the minimal perturbation of an input that will produce the wrong output in the network? One way of avoiding this would be to introduce these types of images in the training set and to repeat the whole process. 24.6 Feature Maps in CNNs One of the most important concepts when working with

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第24章《卷积神经网络》(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/c024b.xhtml) - 结论依据:摘录中关于补充反例后仍需重新验证的定义、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:For complex architectures, adversarial examples are obtained as an optimization problem: What is the minimal perturbation of an input that will produce the wrong output in the network? One way of avoiding this would be to introduce these types of images in the training set and to repeat the whole process. 24.6 Feature Maps in CNNs One of the most important concepts when working with