Ciao to all.
I want to share an idea to the community.
Let’s start with the dog and cat breed example of lesson 1, but instead of having just cats and dogs, imagine to have like 100 base species, and the task would be to identify the breed of each species (each species can have 30 breeds on average).
Or imagine that you have to classify cancer types from different tissue microscope pictures. Each tissue can have different types of cancer.
My guess is that instead of having one single nn to identify all the possibilities, it maybe more effective to train a first level network to distinguish between, for example, cats, dogs and bears, and then train N more networks, one for each category of the first level.
Every further specialization does not need to retrain the other networks, making the process more modular.
This tree organization can be a good way to organize a vision model zoo, ensuring an incremental approach.
Just a stone in the pond to see if anybody else had the same idea or it has been alreasy explored.