The second challenge is related to one-shot learning problem. Usually with face recognition, we have only one photo of each of the persons to recognize. So let's say that we want to recognize employees as they arrive in the morning.

Usually, you really could have just one photo of the employee, or maybe in the best cases, very few of them.

With the knowledge we now have, we can, of course, feed all these photos to a convolution neural network, and through a softmax, could have the number of classes as a number of employees:

Actually, that will not work well, and we could not have really robust prediction. Convolution architecture has so far demonstrated really promising results, but don't forget that training involved thousands of images of one type, and two million images in total. Additionally, this solution will not scale as well.

So, this is what will happen if we have a new employee:

We need to modify the softmax to have 70 outputs. So does that mean that we need to retrain the neural network each time an employee comes in?