Is this the same as a neural network?

No. Here, nodes represent collaborating participants, not neurons or computational layers.

Must all participants connect directly?

No. Sparse graphs can allow indirect influence through paths, depending on the algorithm.

Can the network change over time?

Yes, a system can use changing participation or relationships, but algorithms and analysis must account for that behavior rather than assuming a fixed graph.

AI Glossary / Mathematics and optimization / Intermediate

Federated learning network

Definition

A federated learning network is a representation of the participants and collaboration relationships in a federated learning application. In the Aalto formulation, it is a weighted, undirected graph: nodes represent devices or other learning participants, and edges represent selected relationships between them. Nodes hold local datasets and models. Edge weights express the strength of a relationship used to organize learning, such as similarity between the participants' tasks.

What it means

The network explains who collaborates with whom, not just where computers are physically connected. Two sites might communicate over the same internet infrastructure yet be connected differently in a learning graph because their prediction problems differ. Conversely, an intended collaboration must still have a feasible communication mechanism. The graph is therefore a modeling and design choice. It should be distinguished from the complete deployed system, which also includes software, access controls, scheduling and monitoring. This entry follows the source's undirected graph convention; other federated architectures can have different relationships.

How it works

First identify participants and their local learning tasks. Choose which pairs should exchange relevant information and assign documented edge weights. Collecting those weights in an adjacency matrix provides a compact representation. In a graph-coupled objective, the edges can penalize differences between neighboring models, allowing each node to fit local data while receiving influence from selected peers. During training, messages follow the algorithm's prescribed communication pattern. Graph structure affects which information can propagate, how much communication is needed and where failures can separate collaborating groups. A graph alone does not specify the optimization algorithm.

Three practical examples

Regional forecasting teams can connect sites with comparable demand patterns more strongly than sites serving very different markets.
A sensor-modeling study can compare a chain of neighboring devices with a denser collaboration graph to investigate communication and prediction tradeoffs.
A reliability exercise can remove a participant from a proposed graph and identify which remaining sites lose all collaboration paths.

Think of it this way

Think of a professional advice network. People have their own work, choose relevant colleagues and place different weight on their advice. Being in the same building does not make every advice relationship equally useful. The analogy stops short of the mathematics: an edge weight must have a defined role in the learning procedure, and similarity or influence cannot be inferred reliably from a social metaphor.

Worked example

Consider three participants A, B and C. Let the undirected edge A - B have weight 2, edge B - C weight 1, and omit A - C. In the order A, B, C, the adjacency matrix has rows (0, 2, 0), (2, 0, 1) and (0, 1, 0). It is symmetric. Weighted degrees are 2, 3 and 1. Suppose their scalar model parameters are 1, 2 and 5. A squared disagreement measure, counting each undirected edge once, is 2(1 − 2)² + 1(2 − 5)² = 2 + 9 = 11. This number measures disagreement under the chosen representation; it is not prediction error. Now add edge A - C with weight 0.5. The measure becomes 11 + 0.5(1 − 5)² = 19 for the unchanged parameters. The increase does not show that adding the edge made predictions worse: the objective now evaluates an additional relationship. Originally, removing B disconnects A and C. With the added edge, they remain connected to each other. If an algorithm sends one parameter vector in each direction along every edge per round, the original graph uses four directed transmissions and the expanded graph uses six, before counting retries or other messages. This is a specific communication assumption, not a rule for every federated method.

Illustrative case study

In a fictional deployment study, Linden Sensors groups instruments by the physical conditions they measure. Engineers initially connect every instrument, then discover that two groups have different calibration behavior. They compare a revised graph using meaningful within-group relationships and carefully selected cross-group links. The evaluation records per-device prediction errors and message counts. A cheaper communication pattern is not accepted merely because it has fewer edges; the team also checks whether isolated devices lose useful information. No improvement is assumed in this illustrative case.

Common misconceptions and mistakes

A learning edge does not necessarily mean a dedicated physical cable. A high weight is not automatically a high bandwidth connection; its meaning must be stated. Connectedness allows paths between nodes but does not prove useful statistical collaboration, secure communication or convergence of every algorithm. Adding edges changes both costs and, in some formulations, the learning objective. Zero can represent an absent relationship in an adjacency matrix without implying a cost-free real link. Finally, an undirected similarity graph is a deliberate simplification when influence or communication is asymmetric.

Business use

Document why each relationship exists and how its weight is obtained. Separate task similarity from network capacity if both matter, rather than mixing their units without an explicit rule. Check whether local model parameters are comparable before penalizing their differences. Evaluate the consequences of participant outages and identify nodes that connect otherwise separate groups. Compare graph designs using prediction quality, communication requirements and participation coverage. Revisit the graph when local data distributions change. Avoid assuming that a visually neat network diagram establishes that collaboration benefits every participant.

Check your understanding

In the original three-node example, why does removing B separate A and C?

Answer: B lies on their only path. There is no direct A - C edge, so no collaboration path remains after B and its incident edges are removed.

Related terms

Federated learning system
Edge weight
Generalized total variation
adjacency matrix
graph connectivity
personalized federated learning

Learn next

draw a four-participant chain and a four-participant star. Compare their edge counts and the effect of removing the central participant. Then define what a weight means in one real project and identify evidence that could justify increasing or decreasing it.

Sources

Sources: Alexander Jung, Federated Learning: From Theory to Practice, 2025 author preprint, Section 3.2, https://arxiv.org/pdf/2505.19183 ; Jung and colleagues, The Aalto Dictionary of Machine Learning, https://aaltodictionaryofml.github.io/ADictML_Main.pdf . The three-node matrix, disagreement calculation, transmission count and fictional sensor study are original illustrations under explicitly stated assumptions.

Previous: Federated learning Next: Federated learning system