i am trying to classify some images in classes using the convolutional networks approach. However there are varying numbers of training examples per class. I am worried that that might cause overfitting for the bigger classes. Should i just through all the examples in and see what happens or would it be better to get the minimum number of examples available to the smallest class for each of the big classes discarding at random the rest of their training examples?
Your question is about imbalanced datasets. There is a question about it here and very short two papers here and here. There is a simple method which up-samples the minority class or/and down-samples the majority class, but it may not be very useful because first one replicates similar instances and the second one throws away useful information. I suggest you to augment your classes to equal number of samples if it is possible.
There is no solution to your question. For data science you need make some experiments for your data samples and check changes. It's hard say what will be better for any data. You can test your overfitting in simple way. Try in every iteration compute error for training sample (blue graph) and for test sample (red graph).
You can test it for full data sample and for data sample with removed some features. This trick can help you to see which was better for you.