← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6753 · 生成模型:GPT-6

双向掩码训练不是双向反传特例

理解等级:中级

BERT在输入侧让词位利用左右可见上下文,以选定位置的恢复目标训练。其梯度仍按普通计算图反传,双向描述的是上下文可见性,不是另一种特殊反向传播,也不等于Transformer是增强RNN。

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

      9.3 Bidirectional backpropagation and BERT Sometimes, you want to predict something in the middle of a sequence—perhaps, a masked-out word. Transformers can handle that as well. The model doesn’t need to be limited to reading your text from left to right in a “causal” way; it can read the text from right to left on the other side of the mask as well. When generating

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第09章《可堆叠的Transformer》(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-9.html) - 结论依据:摘录中关于双向掩码训练不是双向反传特例的定义、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:9.3 Bidirectional backpropagation and BERT Sometimes, you want to predict something in the middle of a sequence—perhaps, a masked-out word. Transformers can handle that as well. The model doesn’t need to be limited to reading your text from left to right in a “causal” way; it can read the text from right to left on the other side of the mask as well. When generating