Question

Consider a simple case where we have trained a NN for recognising patterns X₁ to X_n belonging to some classes, say A and B. Suppose we have a new training data X_n+1 that belongs to class C. We want this also to be trained. Will the previous training be impacted? How would it work?

Answer 1

So, according to you the existing classes are A and B. In this case, when you have a new training instance Xn+1 that belongs to class C, it is possible to train this instance as well, without effecting the previous trained model.

Your pre-trained network has a layer, which handles the recognition of 2 original classes. The easiest (and working) trick to introduce the new class, is to use all the layers before the last as granted and add an additional layer (in a new model, or as a parallel one).

Without access to original training data, you would have two options:

1. Freeze all the weights in original layers by allowing "new" model to optimize only new weights. That will give you exactly same predictive power for original 2 classes and might give OK performance for new ones.
2. Train whole network at once (by propagating error of new classes), which might be working for new class(es), but you will end up with ineffective original solution for 2 classes (since weights will be changed for the lower classes and final layer won't be updated to match those changes).