Machine Learning Methods
FedProto outperformed several recent federated learning approaches on multiple datasets in a benchmark tailored for heterogeneous federated learning

In a 2022 paper in the Proceedings of the AAAI Conference on Artificial Intelligence, researchers describe a federated learning method called FedProto. Federated learning is a way to train a shared model across separate clients without pooling their data. In FedProto, clients send abstract class prototypes to a server instead of sending gradients, the update signals that normally change a model's weights.
In benchmark tests designed for heterogeneous federated learning, where clients may differ in data distribution, network latency, input/output space or model architecture, FedProto outperformed several recent federated learning methods on multiple datasets.
Why does heterogeneous federated learning make training harder?
When clients differ in their data, speed or model design, their local gradients may not line up. The abstract states that this heterogeneity usually hinders optimization convergence and generalization performance when knowledge is combined in gradient space.
What did FedProto change?
FedProto exchanges abstract class prototypes instead of gradients. The abstract does not define what a class prototype is. The server aggregates local prototypes from different clients, then sends global prototypes back to all clients. Each client's training minimizes classification error on its local data while keeping its local prototypes close to the global ones. The paper also includes a theoretical analysis of FedProto's convergence rate under non-convex objectives.
Questions about the paper
What is federated learning?
Federated learning trains a model across many clients without collecting their raw data in one place.
What is a class prototype?
The abstract does not define what a class prototype is. Sending prototypes instead of gradients is the main communication change in FedProto.
What did the benchmark show?
On a benchmark setting tailored for heterogeneous federated learning, FedProto outperformed several recent federated learning approaches on multiple datasets.
Limits and what's next
The abstract describes one benchmarking study on multiple datasets, using a benchmark setting designed for heterogeneous federated learning. It does not state which datasets were used or how many clients participated. The abstract reports no independent replication of the result. The convergence-rate analysis is described under non-convex objectives, but the abstract does not state what follow-up work the authors plan. The result is one benchmark finding, and the abstract reports no independent replication.
Source: Yue Tan et al., “FedProto: Federated Prototype Learning across Heterogeneous Clients,” Proceedings of the AAAI Conference on Artificial Intelligence, 2022.