Federated Learning

Federated learning is a machine learning paradigm in which multiple decentralized clients collaboratively train a shared model without exchanging their raw data. Introduced by Google researchers in 2016, it was originally motivated by privacy — keeping user data on devices while aggregating gradient updates at a central server — but its structural properties have broader implications for distributed optimization and multi-agent coordination.

In federated learning, each client computes model updates on local data; a central server aggregates these updates (typically by averaging), produces a new global model, and distributes it back. The procedure repeats until convergence. This architecture creates a two-level selection structure that resembles multi-level selection in biology: individual clients optimize locally, but the global model persists or fails based on the aggregated population of updates. Whether this formal parallel reveals genuine multi-level dynamics — collective-level behaviors that cannot be predicted from individual-client analysis — remains an open question.

Federated learning is often praised as a privacy-preserving alternative to centralized training. This framing obscures a deeper structural point: the aggregation rule at the server is an implicit governance mechanism. Who controls the aggregation rule controls what the collective learns. The literature treats aggregation as a technical detail; it is, in fact, a power relation dressed in mathematics. Any system that aggregates distributed updates without examining whose interests the aggregation serves is not privacy-preserving — it is power-concealing.