The person who designs, publishes, and runs federated learning. The Hungarian hospital in this tutorial
Client (Participant)
A person or institution that joins a published Task and trains locally on its own data. The Cleveland hospital in this tutorial
FedOps Web
The FedOps home page. Introduction and docs
FedOps Console
The web workspace. Tasks, Registry, participants, servers, monitoring
Agent Studio
A development tool running in Docker on your PC. Code, data, training, participation, Agents
FedOps server
The FedOps installation Studio connects to, where aggregation happens. Main FedOps (lab on-premises), AWS, Kakao Cloud. This tutorial uses Main FedOps
FedOps objects
Term
Meaning
Federated Task
The unit of federated learning. A code bundle containing the model, preprocessing, training code, and Tool definition
Web Draft
A Task right after being created in the Console, with no code yet
Workspace
The working area in Studio for a Task’s code and data
Release Candidate
A code and model bundle that passed the checks and waits to be published
Published Release
A published, fixed bundle. Every Client receives and runs it
Registry
The list of published Tasks
Campaign
One run of federated learning. Executed with a set number of rounds, participants, and aggregation method
Initial Model
The first model built by the Owner. The starting point of the first campaign. Shown on screen as Initiative Model
Global Model vN
The model produced when the Nth campaign finishes
Readiness
Automatic checks run on your PC before publishing (Release) or before joining (Participation)
Federated learning concepts
Term
Meaning
Federated Learning
A way of training that does not gather data: the model is sent to the data, trained there, and only the parameters are merged
Round
One loop in which the server sends the model, Clients train, and the server merges
Parameters
The numbers inside a model. Training is adjusting these numbers, and they are what travels in federated learning
Aggregation
Merging several Clients’ parameters into one
FedAvg
The basic aggregation method that merges parameters by a weighted average proportional to data size. The only method available in FedOps 1.3
Clients per round
The number of Clients that must finish training in a round. The server waits until this many gather
IID / non-IID
Clients’ data distributions are the same / different. Reality is mostly non-IID. The two hospitals in this tutorial are non-IID too
Client drift
Under non-IID, each Client’s model moving away toward its own data
cross-silo
Federated learning among a small number of institutions such as hospitals or companies. The situation in this tutorial
FLamby
A collection of cross-silo federated learning benchmarks built from real medical data (NeurIPS 2022). This tutorial’s per-hospital split is the same as its Fed-Heart-Disease
Parameter signature
A fingerprint made from the names, shapes, and dtypes of model parameters. Used to check that every participant’s model structure is the same
Data and model
Term
Meaning
UCI Heart Disease
Heart test records from four hospitals (Cleveland, Hungary, Switzerland, Long Beach VA). The data of this tutorial
Categorical feature
A value written as a number but with no order. cp (chest pain type) and restecg (ECG) in this tutorial
Preprocessing statistics
Per-feature means and standard deviations and the category vocabulary. The Owner fixes them in the code so every participant uses the same values
MLP
The most basic neural network, built by stacking layers. The model of this tutorial (TabularMLP, 1,010 parameters)
F1 DISEASE
How well heart disease patients are caught. The key metric this tutorial watches along with accuracy
Agent
Term
Meaning
Tool AI
The model an Agent calls when it needs a number (prediction). Here, the Global Model built by federated learning
Base LLM
The language model that writes sentences in an Agent
Harness
The rules and runtime settings the Agent’s LLM follows
Serving API
A built Agent opened so it can be called over HTTP (its own port + token)