CoSS: Co-optimizing sensor and sampling rate for data-efficient human activity recognition

초록

Recent advancements in artificial neural networks (ANN) have significantly improved human activity recognition (HAR) using multiple time-series sensors. While employing numerous sensors at high sampling rates typically improves performance, this approach leads to data inefficiency and excessive model complexity, challenging practical deployment on edge devices. To address this, we propose CoSS, a pragmatic framework for the joint optimization of sensor modalities and sampling rates. Central to our approach is a set of trainable parameters, called 'Weight Scores', that quantify the importance of each sensor and sampling rate during a single training phase. These scores guide a pruning process, enabling a data-driven trade-off between classification performance and computational cost. We validate our framework on three public benchmarks: Opportunity, PAMAP2, and MHEALTH. Experimental results demonstrate that configurations selected by CoSS achieve performance comparable to baselines that use all sensors at maximum sampling rates, while drastically reducing resource requirements. For instance, on the MHEALTH dataset, CoSS reduces model size by 62% with a negligible performance decrease of only 0.29%, proving its efficacy in creating efficient and deployment-ready HAR systems.

키워드