Applying the model to unseen data

Question

What are good ways to extend an ML model with a new class without relabeling all previous data?

Problem with data representing classes that weren't present during supervised training

Suppose we have trained a model for classifying a problem with 5 classes. Now if we you use the same model without re-training the model, for a problem with 7 classes what happens? there are still the 5 previous classes with two additional classes.

What conditions the runtime data should have so that the original model perform good on the new data?

What are the terms used by specialists or academics for this problem?

Answer 1

For a fully connected neural network, you plug the two new neurons in the final output layer to all neurons in the previous layer. As weights have not been learnt during training, you can initialize them to some random initializer (Xe or He). During inference time, all neurons in the output layer will have some value, and we usually use a sigmoid activation function in the output layer to condense outputs from all neurons to add up to one.

You'll still have a winner neuron, with the highest value, but might be errorenous as the weights for the two new neurons haven't been tuned to the input data.