Personalized models
We study how shared models can adapt to each system’s data and operating conditions.
Connect learning.Across machines.
For teams with models at multiple installations: coordinate training without exchanging raw operational data. Build your own models with NexuML.
NexuFL Studio / Development workbench
Inspect rounds. Compare experiments.
Local client training, Aggregation, checkpoint continuation and recorded metrics. The orchestration and experiment interface is a development workbench, demonstrated here on synthetic data.
Training rounds & recorded loss — faithful excerpt of the same development UI. Scroll sideways if needed; open full size for the complete view. Inspect focused capture (new tab)
- Track training rounds
- Inspect client contributions
- Compare recorded metrics
NexuFL / coordination architecture
Connect the training workflow.
Bring a model developed in your own ML stack. NexuML is not required. Integration must align the model architecture, update format and local training task across clients.
- 01 / Compatible clients
Local training
Prepare model updates from each participating system.
- 02 / NexuFL coordinator
Coordination & aggregation
Manage rounds, combine updates and record experiment metrics.
- 03 / Client-side checks
Local evaluation
Compare the returned candidate with local baselines before adoption.
Ongoing research / Local adaptation
Adapt learning to each system.
Adaptive aggregation
We study how reinforcement learning can adjust each participant’s contribution as local conditions differ.
Experiments & technical evidence
Inspect the evidence.
Recorded FedAvg experiment
The Studio captures show local synthetic CPU experiments, separate from the university research.
IEEE ICASSP / 2026
Adaptive aggregation research
René Glitza, Luca Becker, Rainer Martin
Peer-reviewed university research on weighting client contributions.