知识卡片

普通读书笔记卡 · 8669.e · 生成模型:GPT-6

域专家学习仍依赖路由

理解等级:高级

以域为单位的专家可为新域单独训练分支,推理时再选择或混合。新域初始化可从相近已有专家出发,但相近度、路由和共享模块影响效果;“分别训练专家”不自动消除所有遗忘或保证未见域泛化。

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

      Leveraging DEMix Layers Transformers can be made more modular by composing the model as a mixture of experts, as shown in Chapter 4. One way to divide the experts is to assign each expert a single domain. This removes the possibility of catastrophic forgetting when learning new domains because each expert is trained separately without affecting other experts. To implement this, Gururangan et al. propose

    参考来源

    - 位置:《Designing Large Language Model Applications》第07章《进阶微调技术》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch07.html) - 结论依据:摘录中关于域专家学习仍依赖路由的定义、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:Leveraging DEMix Layers Transformers can be made more modular by composing the model as a mixture of experts, as shown in Chapter 4. One way to divide the experts is to assign each expert a single domain. This removes the possibility of catastrophic forgetting when learning new domains because each expert is trained separately without affecting other experts. To implement this, Gururangan et al. propose