知识卡片
共享投影控制参数规模
把所有图像标量直接连接到全连接层,会让参数规模随位置数迅速增长。Transformer对每个词元复用查询、键、值投影和逐词元网络,将位置间交互交给动态注意力矩阵。参数不必随词元数平方增加,但标准全局注意力的分数矩阵仍有平方规模。
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
How can we efficiently pass messages across large spatial distances? We already have seen one option: just use a fully connected layer, so that every output neuron after this layer takes input from every neuron on the layer before. However, fully connected layers have a ton of parameters (N2 if their input and output are N-dimensional vectors), and it can take an exorbitant amount of
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:摘录中关于共享投影控制参数规模的定义、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:How can we efficiently pass messages across large spatial distances? We already have seen one option: just use a fully connected layer, so that every output neuron after this layer takes input from every neuron on the layer before. However, fully connected layers have a ton of parameters (N2 if their input and output are N-dimensional vectors), and it can take an exorbitant amount of