I am in the middle of Part 2 and was just thinking about the concept of catastrophic forgetting. In the lecture the solution to catastrophic forgetting is to continue to provide input of the skill you don’t want the model to forget.
Can multiple models be made to work together?
A model can be trained to recognize pictures but can it be trained to recognize other models?
Is it possible for input to travel through a hierarchy of other models to avoid catastrophic forgetting by allowing a model that is most suited for that input to process it? For example, we have a model fine tuned for cats/dogs, airplanes, and cups all in a ‘pool’. A picture of an airplane is passed to this pool via some kind of filter/model and instead of getting passed to cats/dogs or cups - it is passed to airplanes.