Quantum-Assisted Memory-Efficient Training for Parameter-Intensive Wi-Fi-Based Human Activity Recognition
This paper proposes a new quantum-assisted memory-efficient training framework (Q-MET), aimed at improving the efficiency of Human Activity Recognition (HAR) based on Wi-Fi during both training and inference phases. This framework utilizes a hybrid quantum-classical neural network to indirectly generate HAR model parameters, significantly reducing the number of trainable parameters and integrating structured pruning to support deployment on resource-constrained devices. Experimental results show that compared with conventional backpropagation deep learning training, the number of trainable parameters in Q-MET is reduced by 90% to 95%, while maintaining or exceeding classification accuracy; structural pruning achieves a model sparsity of 75% to 85%, with a classification accuracy loss of less than 2%. To the authors’ knowledge, this is the first quantum-assisted method that addresses both training and inference phase memory efficiency issues in HAR systems.