# What is the minimum sample size required to train a Deep Learning model - CNN?

It is true that the sample size depends on the nature of the problem and the architecture implemented. But, on average, what is the typical sample size utilized for training a deep learning framework?

For instance, in a convolutional neural network (CNN) used for a frame-by-frame video processing, is there a rough estimate for the minimum no. of samples required to train the model?

• What kind of data are you trying to recognize? Did you have a certain architecture in mind? You generally see accuracy push up high very quickly with only a few hundred training examples, but the amount needed rises exponentially as you push for higher accuracy. In my experience with CNN's, once you are in the 90% range you may need a number of examples in the thousands. But I imagine it differs a LOT between models and data. – Frobot Feb 8 '16 at 22:07
• @Frobot: To simply put it, given an image, I would like to locate 14 specific $(x,y)$ points in the image at the same time. – Ébe Isaac Feb 9 '16 at 1:07
• This highly depends on your concrete problem and your architecture and can not be generalized. However, as a rule of thumb, I would say e.g. for localizing something in images you will need more than just 2 or 3 images per region where you want to localize something. This is really hard to say but most times you see neurons really start to learn invariants and converge if they have like 10 or more similar training samples for every different input situation. So if you have like 9 regions (left, center, top, ...) I would say with roughly 90-100 images you should see some learning effect kick in – daniel451 Feb 10 '16 at 11:42
• Possible duplicate of How few training examples is too few when training a neural network? – Sycorax Aug 17 '18 at 3:26