Roadmap for the future

Here is a shortlist of future revisions on our roadmap that may make it to minor versions rather than be delayed to the 3.0 release. This roadmap is not engraved; it may and will be updated as development effort go forward.

Users of Declearn and potential contributors are encouraged to voice their opinion, advice and concerns, whether here or via e-mail.

New features

• Implement secure aggregation and deploy it on top of the current federated learning process (either organically or via a new pair or main classes).

• Introduce APIs for learning-rate scheduling, layer-wise optimization rules, and clients sampling rules.

• Implement fairness-aware federated model training algorithms and make them deployable (either as part of or via subclasses of the current orchestration classes).

Revisions to process orchestration

• Revise or replace TrainingManager to provide with training and evaluation routines that are de-coupled from the (un)packing of information in messages.

• Revise configuration tools to make it easy to specify options in a sensible and modular way. Our aim is notably to cut down the need for boilerplate code when adding new options on top of the current federated learning process.

• Introduce an API to define federated routines (perhaps designed as payloads to dedicated Message subclasses), that may be extended and combined into processes in a more modular fashion than the current code allows.

• Revise metadata collection and/or implement new tools for the (generic) sharing and aggregation of federated analytics about users' data.

Revisions to network communications

• Revise serialization schemes, utils and possibly backend technology so as to speed up transforms and lighten up communicated messages.

• Revise the logic for communication, introducing a communication server acting as a central postmaster decoupled from computations' orchestration. Current server / client roles from centralized federated learning would still be made available via (hopefully) API-stable wrappers around new peer endpoints.