statistics - Can a generally trained deep learning classifier be used to classify within subclasses? -

suppose, deep learning classifier trained distinguish between images of cars, ships, trucks, birds, horses , dogs. training data birds yellow birds.

can trained classifier used detect yellow birds within birds image data set ? image data example here. data can other things dna sequences too. please bear me if question non-sensical or basic.

in example mentioned not learning classifier discriminate cars, ships, trucks, birds, horses , dogs between 5 first things mentioned , yellow birds. means when score out birds unit - assuming model performing , dataset sufficiently large - might assume able discriminate between different objects , yellow birds - when these different objects other birds. of course - there small probability learn discriminate among birds , different objects using shapes - it's small in opinion taken account. of course - might check generating appropriate testing dataset.

in general - depends on many factors. 1 of them architecture , design of network. discriminating yellow birds different coloured 1 should easy because of convolution of coloured images nature. in different cases - might not obvious. other thing how far conceptually these classes want discriminate each other. if e.g. example - other class can build out of same concepts learnt 1 - might have problem - cause network might learn them indicators of yellow birds.

so best thing design appropriate testing dataset , perform comparision between scores of different classes. if prove score performs - done. if not - need retrain network.