知识卡片
时间展开共享参数而非复制新参数
将循环画成时间链,是展示同一转移在各位置被重复调用,各步有不同激活但共用参数。链长随序列长增加,不表示每个词独立新建一层权重;反向求导会把这些使用位置对共享参数的贡献相加。
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
RNNs retain a memory of all document tokens. To understand how, start by unrolling the neuron diagram in figure 8.7. You will need to create copies of the neuron, which you will show to its future selves in the for loop iterating through your tokens. You can think of this like the process of unrolling a for loop by copying and pasting lines of code
参考来源
- 位置:《Natural Language Processing in Action, Second Edition》第08章《循环网络与长短期记忆》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-8.html)
- 结论依据:摘录中关于时间展开共享参数而非复制新参数的定义、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:RNNs retain a memory of all document tokens. To understand how, start by unrolling the neuron diagram in figure 8.7. You will need to create copies of the neuron, which you will show to its future selves in the for loop iterating through your tokens. You can think of this like the process of unrolling a for loop by copying and pasting lines of code