← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9261.a · 生成模型:GPT-6

初始化应针对模型和对称性

理解等级:中级

神经网络隐藏单元全相同初始化可能保持对称,使它们学到相同特征,但不能推出任何模型都必须随机初始化。预训练初始化也可能有帮助或负迁移,不能保证一定减少所需训练。

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

      Initialization: the process of assigning parameter values at the beginning of training is called initialization. These parameter values will be adjusted during the training process. After the model is selected, initial parameter values are required and it is often not desirable to set them to all 0’s, 1’s, or a combination. Instead, the initial model parameters should be chosen randomly. It is common practice to

    参考来源

    - 位置:《Machine Learning Fundamentals: Concepts, Models, and Applications》第09章《训练机器学习模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Machine Learning Fundamentals: Concepts, Models, and Applications/_epub-src/OEBPS/xhtml/23_chapter09.xhtml) - 结论依据:摘录中关于初始化应针对模型和对称性的定义、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:Initialization: the process of assigning parameter values at the beginning of training is called initialization. These parameter values will be adjusted during the training process. After the model is selected, initial parameter values are required and it is often not desirable to set them to all 0’s, 1’s, or a combination. Instead, the initial model parameters should be chosen randomly. It is common practice to