Overtrain a CNN?

[Johnny ho]

This may sound silly: will the CNN overtrain? From what I have seen so far, having many epochs and parameters are the key of success of CNN. And there is almost nothing I found is about overtrain a CNN.

However, I have encountered some issue with my dataset. I keep training for a long number epochs and the training accuracy is increasing. However, my validation set accuracy increases at first and then start to decrease after ~20 epochs (i ran for about 30 epochs).

Is this the problem of my dataset? Or I did something wrong? Or I overtrain the CNN?

I implemented my CNN in torch.

[soumith]

Yes, CNNs can get to 100% accuracy easily on a small dataset.

--
You received this message because you are subscribed to the Google Groups "torch7" group.
To unsubscribe from this group and stop receiving emails from it, send an email to torch7+un...@googlegroups.com.
To post to this group, send email to tor...@googlegroups.com.
Visit this group at http://groups.google.com/group/torch7.
For more options, visit https://groups.google.com/d/optout.

[Johnny ho]

THanks for the reply. What do you mean by small? You mean the number of class (e.g. 2) or the number of training example (I have about 100k)?

On Tuesday, November 24, 2015 at 11:10:53 PM UTC-8, smth chntla wrote:

Yes, CNNs can get to 100% accuracy easily on a small dataset.