Can we say that overfitting occur in this plot?

4 次查看（过去 30 天）

显示更早的评论

Ane Kang 2018-8-22

1
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/415780-can-we-say-that-overfitting-occur-in-this-plot

评论： Greg Heath 2018-9-23

Performance plot from ANN is obtained as shown below.Does overfitting happen here?

回答（1 个）

Greg Heath 2018-8-23

1
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/415780-can-we-say-that-overfitting-occur-in-this-plot#answer_333850

在 MATLAB Online 中打开

This is a case of

*OVERTRAINING AN OVERFIT NET*

There are at least 3 ways to avoid this:

 1. *DO NOT OVERFIT:* 
 Make sure that the number of unknown weights, Nw does not 
 exceed the number of training equations, Ntrneq.
 2. *DO NOT OVERTRAIN* 
 In particular, the problem is not necessarily the 
 overfitting. Overfitting is easily mitigated by NOT 
 OVERTRAINING
    a. Use a VALIDATION set to implement "EARLY STOPPING".
    b. Use "REGULARIZATION" via TRAINBR to implement 
       "BAYESIAN RREGULARIZATION"
 See

https://www.mathworks.com/matlabcentral/answers/280818-how-to-solve-overtrained-nn-with-validation-stop

Hope this helps.

Thank you for formally accepting my answer

Greg

And in the first plot (original question), the val error rate only increases for 5 continuous epochs, during epoch 6-11, while the training error is decreasing. Since 5 is less than the default of 6, the first plot shows overtraining. Is this interpretation right?