知识卡片
复现闭包与训练不确定性
模型定义强调「整个训练环境的闭包」,是因为真正复现一个模型需要把大量软件与数据精心、相互一致地做版本管理——这常被忽视,却直接决定复现性。更麻烦的是,ML 尤其分布式部署中的软件带有相当非确定性:网络与并行计算让更新顺序带随机性。因此同一配置、同一数据训练两次,产出的训练模型可能相同也可能显著不同。这意味着模型并非可由规格完全确定的产物,版本、随机种子与训练环境的任何变化都可能改变结果,给定位、回归对比与回滚增加难度。
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
原书出处 · 1 处
- Basic Introduction to Models · 查看原书依据
Note that some of the software that we use in ML, especially in distributed deployments, has a fair bit of nondeterminism. As a result, the exact same configured model trained twice on the exact same data may or may not produce a significantly different trained model. Although in this book we try to not confuse these concepts, be aware that they are not carefully differentiated in the industry as a whole and that we also may b
参考来源
- 位置:《Reliable Machine Learning》第03章《Basic Introduction to Models》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch03.xhtml)
- 结论依据:原文明确阐述该主题,并给出定义、实践或反例。
- 原始内容:Note that some of the software that we use in ML, especially in distributed deployments, has a fair bit of nondeterminism. As a result, the exact same configured model trained twice on the exact same data may or may not produce a significantly different trained model. Although in this book we try to not confuse these concepts, be aware that they are not carefully differentiated in the industry as a whole and that we also may b