Towards Parameter-Efficient Quantum Transformers for Human Activity Recognition

Abstract

Wearable sensors and smartphones have made human activity recognition (HAR) a key component in applications ranging from healthcare monitoring to context-aware services. However, deploying accurate models on resource-constrained devices requires architectures that are not only performant but also parameter- and compute-efficient. In this thesis, we investigate a hybrid quantum–classical transformer-style architecture, called QTModel, for multivariate time series classification with a focus on wearable-sensor HAR. QTModel combines a strong 1D convolutional backbone with stacked quantum feature map blocks implemented in TorchQuantum, integrating quantum-inspired self-attention and a final quantum gate into a compact classification pipeline. We benchmark QTModel against several common baselines, namely: ResNet, EfficientNet, MnasNet, MobileNet, MobileNetV2, and TSLANet. We conduct our experiments across five public HAR datasets: WISDM, UCI HAR, MotionSense, MHEALTH, and PAMAP2. After consistent preprocessing, class balancing by downsampling, and a unified 6-fold cross-validation protocol, we evaluate all models in terms of accuracy, F1-score, precision, recall, and ROC-AUC, and further quantify their parameter counts and FLOPs using ptflops. The experimental results demonstrate that QTModel achieves competitive performance compared to strong mobile CNN and transformer-like baselines while maintaining a favorable trade-off between predictive accuracy and model complexity. These findings suggest that quantum feature maps and quantum-inspired attention blocks are a promising direction for parameter-efficient deep models in HAR and, more broadly, multivariate time series classification.

Date of Award	2025
Original language	American English
Awarding Institution	HBKU College of Science and Engineering