Is a local model always a small model?

No. Local identifies its association with a node, not its size.

Must all nodes use the same hypothesis space?

Not in the supplied general definition, although particular algorithms require compatible models.

Is this the same as local regression?

No. Local regression fits using a neighborhood of observations; the supplied term concerns a model class assigned to a federated-learning node.

AI Glossary / Statistics and probability / Intermediate

Local model

Definition

In the supplied Aalto dictionary's federated-learning terminology, a local model is the hypothesis space assigned to a particular node of a learning network. It specifies the prediction functions that node is allowed to learn. Different nodes may have different local model classes. This precise meaning should be distinguished from a particular fitted predictor, from neighborhood-based local regression, and from the everyday phrase for a language model running on a personal computer.

What it means

A hypothesis space is a set of possible prediction rules. For one node, it might contain straight-line functions; for another, it might contain small neural networks or another suitable family. Training chooses a hypothesis from the available space using data and the learning objective.

In practical federated-learning discussions, people also use local model to mean a client's current fitted parameters. To avoid confusion, this entry distinguishes the local model class from the fitted local predictor. The supplied definition concerns the former, while operational algorithms often exchange updates to the latter.

How it works

Associate each participating node with its data, prediction task, and hypothesis space H(i). Choose a local loss describing errors on that node's data. A collaborative learning objective may couple nodes through shared parameters, penalties on disagreement, exchanged predictions, or another defined mechanism.

If all nodes use a common parameterization, some algorithms can aggregate parameter updates directly. Different model classes may require another compatibility mechanism. The mere fact that models are local does not make averaging their parameter vectors meaningful.

Evaluate both local and collective performance. Data distributions, sample sizes, hardware, and participation can differ across nodes. Keeping raw data at a node can reduce data movement, but transmitted updates may still reveal information unless suitable protections are built into the system.

Three practical examples

A network of factories assigns prediction classes to sites according to available sensors and the process each site must model.
Devices with different computing limits use explicitly supported local model classes within a personalized learning design.
A federated research project distinguishes the shared architecture from each participant's current parameter values, preventing reports from confusing model-class differences with differences caused by training data.

Think of it this way

Think of each workshop receiving a toolbox. The toolbox determines which shapes it can build; the finished object is one particular choice made using that toolbox and local materials. Workshops can have identical toolboxes but produce different objects, or different toolboxes that cannot be combined by simply averaging their contents. The analogy separates a model class from a fitted result.

Worked example

Consider two fictional nodes with scalar input x and numeric outcome y. Node A's local model class contains affine functions a + bx. Node B's class contains quadratic functions a + bx + cx squared. Node A has two coefficients; node B has three.

For a controlled comparison, give both nodes the same three teaching observations: x = -1, 0, and 1, with y = 1, 0, and 1. Least squares within node A's class gives a = 2/3 and b = 0. Its predictions are all 2/3, with squared errors 1/9, 4/9, and 1/9. Mean squared error is 2/9, approximately 0.2222.

Within node B's class, a = 0, b = 0, and c = 1 fits all three observations exactly, giving zero training error. This illustrates a difference in representational capacity, not superior future performance: the larger class could overfit noisy data.

Directly averaging the two coefficient vectors is undefined without specifying how their coordinates correspond. Embedding the smaller model into the larger space is one possible design choice, but would create a particular collaboration rule that still needs justification. The example itself performs no federated training rounds.

Illustrative case study

Fictional case: Seabright Manufacturing wants several sites to learn collaboratively. One site has two temperature sensors, another has additional vibration measurements, and their prediction pipelines were built independently.

The project lead first documents each node's inputs, outputs, and model class. Engineers find that parameter positions do not have common meanings across the systems, so they avoid treating the existing parameter arrays as directly interchangeable.

They choose a compatible collaboration design and evaluate it at every site as well as overall. A gain in the pooled score is not accepted as evidence that every site benefits. The specification keeps model classes, trained predictors, and transmitted information separate.

Common misconceptions and mistakes

Do not assume local means trained in isolation or disconnected from a wider network. A local hypothesis space can participate in a coupled learning objective.

Do not confuse different fitted weights with different hypothesis spaces. Two clients can share the same class but learn different predictors. Conversely, equal-length parameter arrays do not guarantee compatible meanings. Local computation also does not guarantee privacy, fairness, or representativeness. Each of those properties needs its own design and evaluation rather than being inferred from the word local.

Business use

This distinction helps businesses specify distributed AI projects clearly. Contracts and technical plans should state what each participant can learn, which components are shared, and whether personalization is expected.

Assess model capacity against local sample sizes, input availability, and computing limits. Record how updates or predictions are combined and how the system handles absent or incompatible nodes. Measure service quality at individual sites so that an aggregate improvement does not hide deterioration for smaller participants or unusual operating environments.

Check your understanding

Two nodes use the same neural-network architecture but have different fitted weights. Must their local hypothesis spaces differ?

Answer: No. They can choose different predictors from the same hypothesis space. Architecture and constraints define the available class; training determines a particular fitted member.

Related terms

Hypothesis space
Federated learning
Model parameter
local dataset
personalized learning
model parameters
global model
heterogeneous models

Learn next

Next, distinguish shared-model federated learning from personalized or heterogeneous-model collaboration. Then study the compatibility assumptions of an aggregation method before applying it to independently developed predictors.

Federated learning

Sources

Sources: The Aalto Dictionary of Machine Learning, Local Model, exact federated hypothesis-space definition: https://aaltodictionaryofml.github.io/ADictML_Main.pdf . McMahan and colleagues, Communication-Efficient Learning of Deep Networks from Decentralized Data (2017), primary shared-model algorithm and update discussion: https://proceedings.mlr.press/v54/mcmahan17a/mcmahan17a.pdf . Both sources were read. The model-class comparison and fictional case are original; both least-squares fits and errors were independently checked. Scope follows Aalto's specific terminology.

Previous: Local linear trend model Next: Local outlier factor