知识卡片
推理机制假说不能取代泛化测量
将任务分解为图中子任务并匹配训练见过的子图,是解释部分组合任务表现的一种研究视角。它提示可用新组合、难度增长和分布外任务检验泛化,不能凭此宣称所有Transformer只会这一种机制。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 对齐训练与推理
原书出处 · 1 处
- · 查看原书依据
linearized subgraph matching. In this perspective, a task is said to be represented as a directed graph, where the directed graph represents the steps involved in solving the task. The subtasks making up the tasks correspond to subgraphs within the directed graph. Transformers solve subtasks by matching the corresponding subgraph to subgraphs seen during the training data. I recommend reading the entirety of Dziri et
参考来源
- 位置:《Designing Large Language Model Applications》第08章《对齐训练与推理》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch08.html)
- 结论依据:摘录中关于推理机制假说不能取代泛化测量的定义、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:linearized subgraph matching. In this perspective, a task is said to be represented as a directed graph, where the directed graph represents the steps involved in solving the task. The subtasks making up the tasks correspond to subgraphs within the directed graph. Transformers solve subtasks by matching the corresponding subgraph to subgraphs seen during the training data. I recommend reading the entirety of Dziri et