Project overview
CurFed-CHARM addresses a practical deployment problem: event-camera data may be distributed across multiple devices and environments, while privacy or bandwidth constraints prevent raw data centralisation. The framework begins with a client containing all action classes and progressively incorporates additional clients after performance thresholds are reached.
Core contributions
- Curriculum sequential federated learning for progressively integrating heterogeneous clients.
- A channel-hierarchical action-recognition model with attention-enhanced 3D convolutional processing.
- Privacy-preserving training without sharing raw event-camera recordings.
- Specialized metrics for client convergence, knowledge transfer, catastrophic forgetting, and communication efficiency.
Reported evidence
The reported average accuracy exceeds FedAvg at 92.99%, FedProx at 92.42%, and FedCurv at 92.09%. The study also analyses convergence behavior, client progression, forgetting, and communication overhead.
Datasets and evaluation
- E-FedAct, structured to represent heterogeneous client-wise action data.
- Non-IID scenarios with different action-class distributions across clients.
- Evaluation across standard classification measures and federated-learning-specific metrics.
Relevance to human-in-the-loop robotics
Distributed human-aware robots and smart environments cannot always centralize sensitive recordings. CurFed-CHARM provides a route to collaborative model learning across devices while keeping raw event data local and controlling communication demands.