← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6732.c · 生成模型:GPT-6

初始状态与初始参数分别管理

理解等级:中级

初始隐藏激活决定每条序列从何状态开始,参数初始化决定训练前转移函数的权重。复用随机种子不自动让特定隐藏维具有固定可解释职能,也不保证全局最低损失;应分别记录两种初始化及复现条件。

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

      Tip In some cases, it can help to seed your neural networks with an initial hidden state other than all zeros. Jonathan Frankle and Michael Carbin found that being intentional about reuse of good initialization values can be key to helping a network find the global minimum loss achievable for a particular dataset.25 Their approach is to initialize all weights and biases, using a random

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第08章《循环网络与长短期记忆》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-8.html) - 结论依据:摘录中关于初始状态与初始参数分别管理的定义、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:Tip In some cases, it can help to seed your neural networks with an initial hidden state other than all zeros. Jonathan Frankle and Michael Carbin found that being intentional about reuse of good initialization values can be key to helping a network find the global minimum loss achievable for a particular dataset.25 Their approach is to initialize all weights and biases, using a random