知识卡片
归一化汇集不同于计数
若把检测到的动物位置赋权1并求和,可以得到数量;但Softmax使每行权重和为1,对常数1值向量汇集仍得到1。书中的计数直觉需要未归一化权重、携带数量的值表示或额外机制,不能直接用归一化平均复制求和结论。
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
As a concrete example, consider that we want to be able to ask questions about different objects in our safari example image, such as how many animals are in the photo. Then one strategy would be to attend to each token that represents an animal's head, and then just count them up. The f would take as input the text query, and would produce as
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第26章《Transformer》(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/c026.xhtml)
- 结论依据:摘录中关于归一化汇集不同于计数的定义、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:As a concrete example, consider that we want to be able to ask questions about different objects in our safari example image, such as how many animals are in the photo. Then one strategy would be to attend to each token that represents an animal's head, and then just count them up. The f would take as input the text query, and would produce as