Significance of cross-validation in tuning weights in CNN

Hi ML experts,

In MATLAB, the dataset is divided into training and validation data using splitEachLabel. The training data is used to train the CNN model and tune its weights in such a way that the error is minimal.

Can anyone please tell me the whether the inbuilt validation that happens during the CNN training has any role towards fine-tuning the filter weights? Does it make any difference or is it just to see how well our model is generalized before we do the actual testing?

Answer 1

[Prajit T R]

Hi Venkat

The validation set is separate from the training set, and hence they do not influence the filter-weights.

The validation data is used to test the accuracy of the model and help you decide whether you have to change the hyper-parameters or not. Basically, it is a method to determine the effectiveness of the actual model before the actual testing takes place.

This is a recommended practice, because in its absence, the model would be highly sensitive to the training data.

Hope this helps.

Cheers.

[Venkat]

Thanks, Prajit for your input. So the validation set is just to see how well my trained model sees an unknown set of data before I test it with the actual test data. Great.

Cheers

[Greg Heath]

In particular, if the validation performance decreases for 6 (default but adjustable) continuous epochs, training will stop.

Hope this helps.

Greg

[Venkat]

Hi Greg,

So Validation also help in early stopping, correct?

Thanks