知识卡片

普通读书笔记卡 · 9113.c · 生成模型:GPT-6

黑箱机制解释需要实验支撑

理解等级:中级

书中对大型视觉语言模型使用边缘、频率滤波和几何推理的叙述,可作为待验证的机制假说。仅凭它能描述几何图形,不能断定内部确实逐步计算了消失点;要用受控干预、表征或行为实验区分可能解释。

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

      This method is remarkably accurate, and now you know how it works. At one level Gpt4-v is a transformer trained on massive data, using an autoregressive modeling objective (chapter 26). On another level, you can understand what rules that transformer may have learned: it detects edges in its first few layers, uses filters that act like wavelets tiling Fourier space, and it reasons about perspective

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第55章《重访简单视觉系统》(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/c055b.xhtml) - 结论依据:摘录中关于黑箱机制解释需要实验支撑的定义、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:This method is remarkably accurate, and now you know how it works. At one level Gpt4-v is a transformer trained on massive data, using an autoregressive modeling objective (chapter 26). On another level, you can understand what rules that transformer may have learned: it detects edges in its first few layers, uses filters that act like wavelets tiling Fourier space, and it reasons about perspective