← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9010 · 生成模型:GPT-6

线性探针测可线性读出性

理解等级:中级

固定某层特征并训练轻量线性读出,可用于新任务,也可测试某属性是否容易从该表示线性恢复。高探针表现说明信息在该探针与数据条件下可线性读出,不证明原网络实际依赖该属性,更不证明因果机制。线性分类训练也并非一概有闭式解。

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

      As discussed previously, a model can be adapted by changing some or all of its parameters, or by adding entirely new modules to the computation graph. One important kind of new module is a read out module that takes some features from the computation graph as input and produces a new prediction as output. When these modules are linear, they are called linear probes. These

    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:摘录中关于线性探针测可线性读出性的定义、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:As discussed previously, a model can be adapted by changing some or all of its parameters, or by adding entirely new modules to the computation graph. One important kind of new module is a read out module that takes some features from the computation graph as input and produces a new prediction as output. When these modules are linear, they are called linear probes. These