Physics / mechanism
Federated learning (FL) trains a shared model across distributed data sources without moving raw data to a central server. Each node runs local SGD on its own dataset, then transmits only gradient updates or model weight deltas to an aggregator (commonly FedAvg). The aggregator averages updates—optionally weighted by local dataset size—and pushes the revised global model back. Key parameters: communication rounds (typically 50–1000), client participation fraction, local epochs per round, and differential privacy noise budget (ε, δ). State of the art: Google’s production FL on Gboard handles ~500M devices; academic benchmarks show <2% accuracy gap vs. centralized training on IID data, larger gaps on non-IID distributions. Secure aggregation adds cryptographic overhead of ~10–30% latency.
Competitive landscape
Central server training with strict data-sharing contracts is the direct alternative—lower complexity, higher accuracy, but legally untenable under GDPR/HIPAA in many verticals. Split learning partitions the model itself rather than data. Swarm learning removes the central aggregator via blockchain coordination. On-device inference (TinyML) is adjacent but orthogonal—inference, not training.
Companies using
Connected ideas
Sources
Frontier (open questions)
- To be added.