← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6635.c · 生成模型:GPT-6

专家门控不等于人类专业标签

理解等级:中级

门控根据表示给专家分配权重或选择,专家常是前馈子网。训练未必让每个专家对应清楚的人类领域,编号也不表达专业身份;要解释路由行为需独立观察,而非从“专家”一词推断能力。

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

      A key component of the MoE architecture is the gating function. The gating function helps decide which expert is more suited to process a given input. The gating function is implemented as a weight applied to each expert. The experts are typically added to the feedforward component of the Transformer. Therefore, if there are eight experts, then there will be eight feedforward networks instead of

    参考来源

    - 位置:《Designing Large Language Model Applications》第04章《架构与学习目标》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch04.html) - 结论依据:摘录中关于专家门控不等于人类专业标签的定义、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:A key component of the MoE architecture is the gating function. The gating function helps decide which expert is more suited to process a given input. The gating function is implemented as a weight applied to each expert. The experts are typically added to the feedforward component of the Transformer. Therefore, if there are eight experts, then there will be eight feedforward networks instead of