ML笔记:Classification: Probabilistic Generative Model

时间:2022-10-16 18:09:42

用回归来做分类:

远大于1的点对于回归来说就是个error,

为了让这些点更接近1,会得到紫色线.

可见,回归中定义模型好坏的方式不适用于分类中.---回归会惩罚那些太过正确的点

ML笔记:Classification: Probabilistic Generative ModelML笔记:Classification: Probabilistic Generative Model

如何计算未出现在训练数据中的点属于某类的概率?

假设该类对应的训练数据采样于一个高斯分布.

可以用该训练数据来估计该高斯分布的参数.

ML笔记:Classification: Probabilistic Generative Model

基本思路:

很多不同参数的高斯分布都可以采样出训练数据,但是可能性不同,

选出其中可能性最大的那个高斯分布对应的参数.---最大似然估计

ML笔记:Classification: Probabilistic Generative ModelML笔记:Classification: Probabilistic Generative ModelML笔记:Classification: Probabilistic Generative Model

ML笔记:Classification: Probabilistic Generative ModelML笔记:Classification: Probabilistic Generative ModelML笔记:Classification: Probabilistic Generative ModelML笔记:Classification: Probabilistic Generative ModelML笔记:Classification: Probabilistic Generative ModelML笔记:Classification: Probabilistic Generative ModelML笔记:Classification: Probabilistic Generative ModelML笔记:Classification: Probabilistic Generative ModelML笔记:Classification: Probabilistic Generative ModelML笔记:Classification: Probabilistic Generative Model

假设sigma相同时,可以得到线性函数.

该生成模型是通过计算miu1,miu2,sigma来间接计算线性函数参数w,b的,

可以通过判别模型直接计算,如Logistic Regression.

ML笔记:Classification: Probabilistic Generative Model