知识卡片

普通读书笔记卡 · 9011.a · 生成模型:GPT-6

学习提示优化共享输入参数

理解等级:高级

像素提示可写为 min_p Σ_i L(f(x_i+p),y_i),冻结 f,针对目标任务训练共享 p。它与优化对抗扰动形式相似,目的和目标方向却不同;不是对每个样本都单独求一份任意提示。输出类别接口不同还可能需要新投影或读出头。

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

      However, here we are not adding noise that hurts performance but instead adding noise that improves performance on a new task of interest. Given a set of M training examples that we would like to adapt to, , we learn a prompt via the following optimization problem:

    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:摘录中关于学习提示优化共享输入参数的定义、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:However, here we are not adding noise that hurts performance but instead adding noise that improves performance on a new task of interest. Given a set of M training examples that we would like to adapt to, , we learn a prompt via the following optimization problem: