知识卡片
位置表示补充序列顺序线索
单纯无位置信息的全可见自注意力不能自动区分所有排列关系,位置机制与遮罩等设计提供顺序线索。绝对向量、注意力偏置、旋转编码采用不同注入方式,应连同架构和训练范围说明。
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
- · 架构与学习目标
原书出处 · 1 处
- · 查看原书依据
Positional Encoding As discussed earlier, pre-Transformer architectures like LSTM were sequence models, with tokens being processed one after the other. Thus the positional information about the tokens, i.e., the relative positions of the tokens in a sequence, was implicitly baked into the model. However, for Transformers all calculations are done in parallel, and positional information should be fed to the model explicitly. Several methods have
参考来源
- 位置:《Designing Large Language Model Applications》第04章《架构与学习目标》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch04.html)
- 结论依据:摘录中关于位置表示补充序列顺序线索的定义、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:Positional Encoding As discussed earlier, pre-Transformer architectures like LSTM were sequence models, with tokens being processed one after the other. Thus the positional information about the tokens, i.e., the relative positions of the tokens in a sequence, was implicitly baked into the model. However, for Transformers all calculations are done in parallel, and positional information should be fed to the model explicitly. Several methods have