知识卡片
学习位置编码需要检验外推
绝对位置向量也可以通过训练学习,让模型选择适合任务的坐标表示。但训练过的网格和长度不保证覆盖新的分辨率或序列长度;改变布局可能需要插值或重新设计。位置编码可用于卷积网络,代价是相应空间对称性也可能改变。
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
Another strategy is to simply let the positional codes be learned by the model, which could potentially result in a better representation of space than sinusoidal codes [109]. While positional encoding is useful and common in transformers, it is not specific to this architecture. The same kind of encodings can be useful for CNNs as well, as a way to make convolutional filters that are
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:摘录中关于学习位置编码需要检验外推的定义、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:Another strategy is to simply let the positional codes be learned by the model, which could potentially result in a better representation of space than sinusoidal codes [109]. While positional encoding is useful and common in transformers, it is not specific to this architecture. The same kind of encodings can be useful for CNNs as well, as a way to make convolutional filters that are