Uncertainty-Aware Federated Learning for Infant Movement Analysis
The paper reports a federated setup for infant movement assessment that nearly matches centralized training without pooling clinical data.
Edmond S. L. Ho frames the work around fidgety movement classification using video-derived skeletal motion data. The method adds MC Dropout to estimate predictive uncertainty, then uses that uncertainty in UA-FedAvg to weight client contributions. In a three-client, cross-subject evaluation, federated learning beat separately trained local models, and the uncertainty-aware variants generally outperformed standard FedAvg. Accepted at IEEE FLTA26. ArXiv · AI/CL/LG's note
Edmond S. L. Ho frames the work around fidgety movement classification using video-derived skeletal motion data. The method adds MC Dropout to estimate predictive uncertainty, then uses that uncertainty in UA-FedAvg to weight client contributions. In a three-client, cross-subject evaluation, federated learning beat separately trained local models, and the uncertainty-aware variants generally outperformed standard FedAvg. Accepted at IEEE FLTA26. ArXiv · AI/CL/LG's note
score 3