NVIDIA's Auto-FL-Research auto-tunes federated learning algorithms
Coding agents propose training recipes, boosting accuracy on 9 of 11 tasks...
NVIDIA researchers (Holger Roth, Ziyue Xu, and team) introduced Auto-FL-Research (AFR), an agentic workflow that automates the search for better federated learning algorithms. AFR employs constrained coding agents that propose and implement changes to key training ingredients—server aggregation rules, client update schedules, local objectives, and model variants—while respecting fixed compute budgets and communication contracts. This allows systematic exploration of the vast design space that typically requires expensive manual tuning.
Evaluated across five healthcare cross-silo tasks (FLamby benchmarks) and six grouped-client LEAF datasets, AFR found algorithmic gains on four out of five healthcare tasks and five out of six LEAF tasks (five-seed repeat evaluations). However, the paper honestly reports mixed outcomes: some gains were due to genuine FL recipe changes, others were recovered by fixed-surface scalar tuning, and some failed under held-out evaluation. This transparency helps identify robust algorithmic improvements versus single-run artifacts, moving toward reproducible FL research.
- AFR uses agentic AI to automatically propose and test FL algorithm variants including optimizer, aggregation, and scheduling changes.
- Outperformed baselines on 4 of 5 healthcare FLamby tasks and 5 of 6 LEAF datasets with 5-seed repeated evaluations.
- The system distinguishes between genuine FL improvements, fixed-surface tuning effects, and seed-sensitive artifacts, improving reproducibility.
Why It Matters
Automates tedious FL hyperparameter search, accelerating discovery of better federated training recipes for privacy-sensitive domains like healthcare.