← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9008 · 生成模型:GPT-6

少样本依赖已有知识

理解等级:中级

从零训练的数据需求,与利用先验适配新任务的数据需求不同。迁移学习先获得知识,再将其用于新问题;一次或少量示例就能适配,不意味着整个系统只看过这些数据。报告样本效率需同时说明预训练来源、任务关系和适配所需标签。

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

      Humans can do this because we have extensive prior knowledge we bring to bear to accelerate the learning of new concepts. Deep nets are data hungry when they are trained from scratch. But they can actually be rather data efficient if we give them appropriate prior knowledge and the means to use their priors to accelerate learning. Transfer learning deals with how to use prior

    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:摘录中关于少样本依赖已有知识的定义、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:Humans can do this because we have extensive prior knowledge we bring to bear to accelerate the learning of new concepts. Deep nets are data hungry when they are trained from scratch. But they can actually be rather data efficient if we give them appropriate prior knowledge and the means to use their priors to accelerate learning. Transfer learning deals with how to use prior