I have some exposure to federated learning and continual learning which are non-iid learning instances. I was wondering can we state the following:
Article Fast Federated Learning by Balancing Communication Trade-Offs
Federated learning is when the dataset is distributed in a non-iid manner over space (different edge-devices at different geographical locations). Meanwhile, continual/incremental learning is when the dataset is distributed in a non-iid manner over time (each time a class is seen). Then, we could state that federated learning and continual learning are subclasses of non-iid learning.
Do you agree with this conclusion? What are the implications in your opinion?
This is an interesting way to put it. And the implications are that any scheme used in federated learning can possibly transfer to the other members of the family: continual learning, non-IID learning.
Dear Milad Khademi Nori ,
Look the link, maybe useful.
https://arxiv.org/pdf/2003.03196.pdf
Regards,
Shafagat
Thanks Jason Gabriel for pointing out the implications of this similarity.
Thanks Shafagat Mahmudova for the useful link to a paper considering the joint scenario where catastrophic forgetting is tackled with regularization and sparsity.
Non-iid machine learning is a challenging task and it's still an open problem. Anyone who can make a dent in it, contributes a lot to machine learning research.
- Badges
- Science topic
- Similar topics
- Psychology
- Mental Processes
- Learning