Enhancing Efficiency in HAR Models: NAS Meets Pruning

Abstract

Real-time monitoring of human activities using wearable devices often requires the deployment of machine learning models on resource-constrained edge devices. Stateof- the-art Human Activity Recognition neural network models, while effective, already suffer from excessive size and complexity. Furthermore, our systematic analysis reveals that, even worse, the computational cost and model size of most SOTA HAR models escalate significantly with an increase in sensor channels. With advancements in sensor technology making it easier to scale sensor deployments that capture human activities, addressing this challenge becomes critical for practical applicability. In this work, we propose an integrated neural architecture search framework to further lighten HAR models. To counteract the negative effects of scaling at deployment time, the proposed framework simultaneously selects and reduces the number of sensor channels, prunes filters, and decreases the temporal dimensions while training the model on optimized hardware. This results in models that are not only smaller in size but also have less model complexity. We conducted experiments on three HAR datasets, demonstrating that our framework outperforms two state-of-theart pruning methods in reducing model size and complexity, while achieving superior performance. Furthermore, we successfully applied our proposed framework to the deployment of a HAR model on a microcontroller, highlighting its feasibility for realworld implementation.

黄逸然

Academic Associates

My research interests include Data Mining, XAI and Human Activity Recognition.