Why is there so much difference in my cnn regression between loss of train and loss of validation?

2 次查看（过去 30 天）

aybike pirol elmas 2020-7-7

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/561173-why-is-there-so-much-difference-in-my-cnn-regression-between-loss-of-train-and-loss-of-validation

编辑： aybike pirol elmas 2020-7-7

I'm doing position prediction with 1D deep learning using EMG signals. In other words, if I say angle estimation, I would have expressed it better. There are signs and angle information I get from 2 simultaneous muscles. In other words, my inputs are 2 channels and the target is angle values.

As with all classical deep learning methods (for regression), I prepared my data first ( like normalization, reshape). Then I give it to the deep learning network I created and provide the training.

here is my cnn model:

But in the results of the train and test (i.e. when I follow the training process: 'Plots', 'training-progress') I encounter something like this: There is a lot of difference between training loss and validation loss.

I can say that network has always learned the train data. So it memorized.

does not look like test data at all.

I felt it was overfitting. And I tried the following:

1-I tried with sgdm and adam and it didn't make much difference

2- I reduced the learning rate.(current learning rate 1e-2)

3- I reduced the size of the data, and I made a downsample to avoid over-memorizing it (because the angles had the following values: for example, 178.50 178.52 178.52 177.88 177.87 )

When the size decreased, yes difference between training loss and validation loss decreased. .still there is difference.it didn't solve my problem. I can't find where I'm mistake.

If anyone could give some insight on this I would greatly appreciate it.

Thank you.