orthogonalization/ one metric

train、dev/test 划分

 开发集和测试集一定来自同一分布  onthe  same distribution

Human level performance & bayes error

human level performance 接近 bayes error,training error 和 bayes 的差距 : bias(avoidable error),dev/test error: variance,决定先调优bias/variance。

structure machine learning projects 课程笔记

  structure machine learning projects 课程笔记

Bias & variance

  Bias: Training error 和 human performance 相差大  (avoidable error

   structure machine learning projects 课程笔记

  variance: training error 和 dev error 相差大

    structure machine learning projects 课程笔记

Error analysis

  structure machine learning projects 课程笔记

  structure machine learning projects 课程笔记

Mismatched dev/test & training set

option 1:讲所有数据集随机打乱,再划分

structure machine learning projects 课程笔记

option 2:dev/test 来源于target dataset,一部分target set 和其他补充数据集作为training set

structure machine learning projects 课程笔记

structure machine learning projects 课程笔记

Bias & variance & data mismatch

  structure machine learning projects 课程笔记

  human performance 和 training error 评估 avoidable error,traing error和 training - dev error 评估 variance,training - dev error 和 dev error 评估 data mismatch,dev error 和 test error 评估 overfit

 structure machine learning projects 课程笔记

deal with mismatch

  structure machine learning projects 课程笔记

      structure machine learning projects 课程笔记

structure machine learning projects 课程笔记

 

迁移学习

structure machine learning projects 课程笔记

end to end

  免去中间组件设计和先验知识,直接学习输入到输出的映射。

  需要大量数据

  structure machine learning projects 课程笔记

相关文章: