← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8941.b · 生成模型:GPT-6

逐词元网络复用同一函数

理解等级:中级

逐词元网络把每一行d维词元分别送入同一个非线性映射F,输出维度通常恢复为d,便于残差相加。它可在特征轴上扩张再压缩,但本身不读取其他行;跨位置依赖来自此前的混合层。书中将F维度写为N的地方应按特征维理解。

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

      Linear combinations only let us linearly mix and recombine tokens, and stacking linear functions can only result in another linear function. In standard neural nets, we ran into the same problem with fully-connected and convolutional layers, which, on their own, are incapable of modeling nonlinear functions. To get around this limitation, we added pointwise nonlinearities to our neural nets. These are functions that apply a

    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:摘录中关于逐词元网络复用同一函数的定义、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:Linear combinations only let us linearly mix and recombine tokens, and stacking linear functions can only result in another linear function. In standard neural nets, we ran into the same problem with fully-connected and convolutional layers, which, on their own, are incapable of modeling nonlinear functions. To get around this limitation, we added pointwise nonlinearities to our neural nets. These are functions that apply a