← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8940 · 生成模型:GPT-6

注意力提供远距离交互

理解等级:中级

局部卷积需要多层传播才能连接相距很远的图像区域。注意力允许一个词元直接汇集其他位置的信息,交互强度由当前内容决定。它扩大信息交换范围,但不保证每次都选择正确对象,也不意味着所有任务都需要全局连接。

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

      CNNs are built around the idea of locality: different local regions of an image can safely be processed independently. This is what allows us to use filters with small kernels. However, very often, there is global information that needs to be shared across all receptive fields in an image. Convolutional layers are not well-suited to globalizing information since the only way they can do so

    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:摘录中关于注意力提供远距离交互的定义、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:CNNs are built around the idea of locality: different local regions of an image can safely be processed independently. This is what allows us to use filters with small kernels. However, very often, there is global information that needs to be shared across all receptive fields in an image. Convolutional layers are not well-suited to globalizing information since the only way they can do so