← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8970.c · 生成模型:GPT-6

预任务服务于编码器训练

理解等级:中级

预测式表示学习先用f将x编码为z,再由g完成预任务。训练结束后可保留f用于其他任务,而不把预任务头的性能当作唯一目标。组合方向是g∘f;网络能解决预任务,并不自动意味着中间表示适合任意迁移任务。

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

      We have already seen many kinds of predictive learning, indeed almost any function can be thought of as making a prediction. In predictive representation learning, the goal is not to make predictions per se, but to use prediction as a way to train a useful representation. The way to do this is first encode the data into an embedding z, then map from the embedding

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第30章《表示学习》(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/c030b.xhtml) - 结论依据:摘录中关于预任务服务于编码器训练的定义、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:We have already seen many kinds of predictive learning, indeed almost any function can be thought of as making a prediction. In predictive representation learning, the goal is not to make predictions per se, but to use prediction as a way to train a useful representation. The way to do this is first encode the data into an embedding z, then map from the embedding