DART-FL: Burst-Aware Multitask Federated Learning under Dynamic Inference Demand at the Edge
DART-FL is a multitask federated learning framework that jointly adapts inference-training resource splitting and task-level training emphasis under dynamic edge inference demand The framework uses an SLO-aware scheduler that reserves minimum resources for inference based on backlog and service capacity, then distributes remaining compute across tasks via a queue-aware DPP-inspired scheduler Dynamic loss weights are mapped from task allocations, prioritizing high-demand tasks in earlier communic
Analysis
TL;DR
- DART-FL is a multitask federated learning framework that jointly adapts inference-training resource splitting and task-level training emphasis under dynamic edge inference demand
- The framework uses an SLO-aware scheduler that reserves minimum resources for inference based on backlog and service capacity, then distributes remaining compute across tasks via a queue-aware DPP-inspired scheduler
- Dynamic loss weights are mapped from task allocations, prioritizing high-demand tasks in earlier communication rounds while maintaining shared backbone + task-specific head architecture
- Evaluated on Stanford Cars and Oxford Flowers 102 with synthetic and real Alibaba trace-derived workloads, showing improved accuracy during burst periods without sacrificing long-term multitask performance
- The approach shifts learning progress of high-demand tasks toward their burst periods, addressing the coupled challenge of coexisting training and inference on resource-constrained edge devices
Why It Matters
This work addresses a critical gap in edge AI deployment where devices must simultaneously handle online inference and federated training under fluctuating demand—common in real-world IoT and mobile scenarios. By dynamically reallocating resources based on inference backlogs and task demand patterns, DART-FL offers a practical pathway to maintain service-level objectives while still achieving competitive multitask learning outcomes, making it highly relevant for practitioners building edge intelligence systems.
Technical Details
- Architecture: Clients train a shared backbone with task-specific heads; the complete multitask model is aggregated through FedAvg across the federated network
- Resource Scheduling: At each interval, DART-FL computes minimum inference resource allocation from backlog and profiled service capacity, then applies a queue-aware DPP (Determinantal Point Process)-inspired scheduler to distribute remaining training capacity across tasks
- Dynamic Loss Weighting: Task-level resource allocations are mapped to dynamic loss weights, giving greater training emphasis to tasks experiencing higher inference demand during earlier communication rounds
- Evaluation: Benchmarked on Stanford Cars and Oxford Flowers 102 datasets under both synthetic and real Alibaba trace-derived workloads, measuring accuracy during burst periods and long-term multitask performance
- Core Innovation: Joint optimization of inference-training resource split and task-level training emphasis, rather than treating them independently
Industry Insight
- Edge device manufacturers and cloud providers should consider demand-aware resource scheduling as a first-class design concern when deploying multitask federated learning, as static resource splits will increasingly fail under variable real-world workloads
- The DPP-inspired scheduling approach could be adapted beyond federated learning to any multi-tenant edge computing scenario where inference demand fluctuates unpredictably across services
- Organizations deploying edge AI should invest in inference demand profiling and backlog monitoring infrastructure early, as these signals are critical inputs for dynamic resource allocation frameworks like DART-FL to function effectively
Disclaimer: The above content is generated by AI and is for reference only.