If you are aiming to predict a number of classes that may be too high for the available data, is there any merit to training multiple models to predict a subset of the classes, each, and combining the outputs?
In particular, this is for an image segmentation problem with ~50 classes and only a few hundred training exams. Intuitively, it seems like a single 50-class model should perform no worse than 2 models with 25 classes each, since the predictions for each class can improve based off of knowing the other classes. But, each individual task of predicting 25 classes vs background is easier to train, and the class to exam ratio is pretty high, so easier training is a need.
Is there any literature or known best practice on this?
