-
Notifications
You must be signed in to change notification settings - Fork 1.4k
Open
Description
我看到你LR里面有附带一个链接,解释为什么classification 的问题loss 用 cross entropy而不是用MSE.
链接里面的解释是因为MSE会有很多局部极小值,曲线不光滑(MSE求导不是一个线性函数吗,为什么会不光滑)
这个问题我也研究了很久,我觉得最好的解释是regression 和 classification 的假设是不同的
regression 的假设是派P(y|x) 服从正太分布,classification 的假设是P(y|x)服从伯努利分布。
具体我自己有写一个推导过程https://github.com/pluszeroplus/Deep-Learning/blob/master/loss/why%20not%20MSE.pdf
LitterWindwind, junqiangchen, jufengjia and Real-UtopiaNo
Metadata
Metadata
Assignees
Labels
No labels