知识卡片
软输出靠近顶点不必精确到达
二分类softmax输出位于一维概率单纯形,训练可使各类点靠近对应独热顶点。有限实数logits通常给严格正概率,未必精确等于0或1;正确argmax分类也不要求达到顶点。几何类比不能把置信度极高等同标签必然正确。
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
原书出处 · 1 处
- · 查看原书依据
we can be more precise. The linear layers perform an affine transformation of the space, while the relus project all negative values to the boundaries of the cone whose dimensions are strictly positive. Layer by layer, over the course of training, the net learns to disentangle these two classes and pull the points toward vertices of the one-simplex, achieving a correct classification of the points!
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第13章《作为分布变换器的神经网络》(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/c013.xhtml)
- 结论依据:摘录中关于软输出靠近顶点不必精确到达的定义、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 be more precise. The linear layers perform an affine transformation of the space, while the relus project all negative values to the boundaries of the cone whose dimensions are strictly positive. Layer by layer, over the course of training, the net learns to disentangle these two classes and pull the points toward vertices of the one-simplex, achieving a correct classification of the points!