DAF: Distillation, Augmentation and Filtering based Framework for Efficient Smartphone Human Activity Recognition

Larger, sophisticated sequential models excel in Human Activity Recognition (HAR) using multivariate time-series data but may not suit compute-constrained smartphones due to latency issues. Knowledge distillation offers a solution by training smaller models based on larger teachers, but a single teacher often struggles to perform uniformly well across diverse activity classes. To address this limitation, we propose the Distillation, Augmentation, and Filtering (DAF) framework, leveraging Multiple-Architecture based Multi-Teacher Distillation (MAMTD). This approach identifies the best-performing teacher model for each activity class and uses Contrastive loss-based Distillation to align a smaller student model with the most effective teachers while distancing it from less effective ones. For challenging categories, a peer student model is employed with data augmentation to focus on areas where the first student struggles. Finally, a novel checkpoint ensemble via probability filtering combines the strengths of both student models, achieving a 21.4-24.6\% increase in accuracy for certain confusing categories compared to typical distilled networks, while maintaining low latency.