Federated Learning
A training approach in which a model is trained across many separate devices or organizations without their raw data ever being centralized, introduced in its modern form by McMahan et al. (Google, 2017). Federated training does not remove the row-level unlearning problem; it adds a coordination challenge to it, since removing one participant's influence may depend on that participant's continued cooperation.
Defined in 3 GAGE programs, which carry 5 distinct definitions of it. The wording above is taught in AI Data Governance: The Data Chair.
How each discipline defines it
The same term does different work depending on who is using it. These are the definitions as each program teaches them, unedited.
A training approach in which a model is trained across many separate devices or organizations without their raw data ever being centralized, introduced in its modern form by McMahan et al. (Google, 2017). Federated training does not remove the row-level unlearning problem; it adds a coordination challenge to it, since removing one participant's influence may depend on that participant's continued cooperation.
A model-training technique in which each device (each robot in a fleet) computes a local model update from its own data, and only that update, not the underlying raw data, is sent to a central server and combined across the fleet, allowing a shared model to improve without raw sensor data ever leaving individual devices.
Training where the model goes to the data on each device and only model updates return, keeping raw data local.
A model-training architecture in which a shared model is trained across multiple organizations' or devices' data without the raw data ever leaving its original source; reduces certain centralization and transfer risks but does not by itself resolve either the memorization test or the provenance test for the resulting model.
A machine learning architecture in which a model is trained collaboratively across many participants' own devices, with each participant's raw data staying local and only model updates shared and aggregated centrally, designed primarily to protect data confidentiality rather than to guarantee the integrity of any individual participant's contribution.
Where it is taught
The exact lessons this term appears in. The first 7 topics of every program are free with a free account.
- Privacy and Data Rights in AI Systems · Ethical and Responsible AI and Operational Governance, AI Literacy & Professional Conduct
- Synthetic data: when it protects, when it launders, and how to tell · Feeding the Machines, AI Data Governance: The Data Chair
- Data poisoning: how an attacker teaches your model on purpose · Poison, Leaks, and the Adversary, AI Data Governance: The Data Chair
- Machine unlearning: the emerging answer to deleting what a model learned · The Frontier Discipline, AI Data Governance: The Data Chair
Terms it appears with
Not an alphabetical neighbourhood: these are the terms taught in the same lessons, ranked by how often they appear together.