知识卡片

普通读书笔记卡 · 9011.c · 生成模型:GPT-6

提示token经注意力作用

理解等级:中级

把可学习 d 维 token 加到图像 patch token 中,注意力便可让它影响其他 token 的表示。训练时可只更新提示向量,冻结原模型参数;这仍可能需要通过模型反传到提示。插入位置、token 数与读出方式属于设计选择,不是只要加 token 就保证适配有效。

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

      The token is a d-dimensional code vector and this vector is optimized to change the behavior of the transformer to achieve a desired result, such as increased performance on a new dataset or task. The attention mechanism works as usual except now all the input tokens also attend to the prompt. The prompt thereby modulates the transformation applied to the input tokens on the first

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第37章《迁移学习与适配》(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/c037.xhtml) - 结论依据:摘录中关于提示token经注意力作用的定义、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:The token is a d-dimensional code vector and this vector is optimized to change the behavior of the transformer to achieve a desired result, such as increased performance on a new dataset or task. The attention mechanism works as usual except now all the input tokens also attend to the prompt. The prompt thereby modulates the transformation applied to the input tokens on the first