I have a rather interesting problem here; I work in the field of image classification for quality assurance. For this I have a dataset of about 1 million images, which I have used to train different defect classes. Now one of these defect types has additional properties (new features of an image class), I would like to teach these new features to the previously trained network, and without re-training the whole previous dataset.
In short: new features of an image class should be taught without affecting the performance of the network on the previous training set too much. Is this possible, and if so, what are some strategies for doing this?
Thanks in advance!