For example, initially, there are 10 classes / 10 output neuron in the model and I want to train them for 50 epochs. When I complete training for 50 epochs, I want to add one extra class/output neuron making the total class and output neuron count as 11 and then train for 50 more epochs. This is different from this post because this deals with the output layer neuron instead of the hidden layer.
The difference between a hidden and output layer should not make the suggested approaches invalid, should it?
I think you could still add the parameter during the training. However, I’m currently unsure about dealing with the optimizer. E.g. would you like to reinitialize the optimizer after changing the layer?