I agree with you, that’s also my intuition (decentralized/local learning should outperform localized). But we also worked for years on multi-agent DL, hoping it will have advantages over centralized (gradients flowing up to the objective function) and it still seems centralized

→ View original post on X — @marek_rosa