A Method for Workout Video Classification via Explainable and Federated Learning.

Area of Science:

• Computer Science
• Machine Learning
• Human-Computer Interaction

Background:

• Wearable devices and smartphones facilitate large-scale human activity data collection for fitness monitoring.
• Centralized storage of workout videos raises significant privacy concerns due to identifiable individuals.
• Federated Machine Learning (FML) offers a privacy-preserving alternative by training models locally on distributed data.

Purpose of the Study:

• To propose and evaluate a Federated Machine Learning approach for workout video classification.
• To enhance the proposed FML method with explainability using Gradient-weighted Class-Activation Mapping (Grad-CAM).
• To assess the impact of different federated configurations on classification accuracy and model interpretability.

Main Methods:

• Developed a workout video classification model using Federated Machine Learning.
• Integrated Gradient-weighted Class-Activation Mapping (Grad-CAM) for explainability.
• Evaluated the approach on a multi-class exercise video dataset across various federated settings (clients, aggregation strategies, communication rounds).

Main Results:

• Different FML aggregation strategies yielded comparable overall accuracy in workout classification.
• Grad-CAM effectively highlighted discriminative regions for exercise recognition.
• Explainability revealed differences in model behavior across aggregation strategies and identified contextual biases leading to misclassifications.

Conclusions:

• The proposed Federated Machine Learning approach with Grad-CAM-based explainability is trustworthy for privacy-preserving workout video classification.
• Explainability enhances understanding of model behavior and potential biases in federated learning for fitness applications.
• The method demonstrates the potential for secure and interpretable AI in personalized fitness monitoring.