知识卡片
训练曲线波动不等于没有学习
随机小批更新、正则与随机层可使训练或验证损失短期升高。应按目标与适当窗口诊断趋势,不能要求两条曲线每一步都单调下降;训练改进而验证恶化则提示选择与泛化问题。
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
- · 训练机器学习模型
原书出处 · 1 处
- · 查看原书依据
Loss computation: a loss function is used to compute the error between the model’s predictions and the ground truth values (a more detailed discussion of loss functions is presented below). In this training stage, the loss (error) is computed on the training set and the validation set. Since the validation set is not used directly to update the model parameters over iterations, it provides an
参考来源
- 位置:《Machine Learning Fundamentals: Concepts, Models, and Applications》第09章《训练机器学习模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Machine Learning Fundamentals: Concepts, Models, and Applications/_epub-src/OEBPS/xhtml/23_chapter09.xhtml)
- 结论依据:摘录中关于训练曲线波动不等于没有学习的定义、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:Loss computation: a loss function is used to compute the error between the model’s predictions and the ground truth values (a more detailed discussion of loss functions is presented below). In this training stage, the loss (error) is computed on the training set and the validation set. Since the validation set is not used directly to update the model parameters over iterations, it provides an