Quantum-Assisted Memory-Efficient Training for Parameter-Intensive Wi-Fi-Based Human Activity Recognition

To Truong An; Jie Zhang; Guolin Yin; Junqing Zhang; Yanjiao Li; Trung Q. Duong; Simon L. Cotton

cs.LGSep 2, 2026

Quantum-Assisted Memory-Efficient Training for Parameter-Intensive Wi-Fi-Based Human Activity Recognition

Authors: To Truong An, Jie Zhang, Guolin Yin, Junqing Zhang, Yanjiao Li, Trung Q. Duong, Simon L. Cotton

Organizations: Centre for Wireless Innovation, School of Electronics, Electrical Engineering and Computer Science, Queen’s University Belfast, Belfast BT7 1NN, UK · School of Computer Science and Informatics, University of Liverpool, Liverpool, L69 3DR, U.K. · Institute of Engineering Technology, University of Science and Technology Beijing, Beijing 100083, China · Faculty of Engineering and Applied Science, Memorial University, St. John’s, NL A1C 5S7, Canada

Abstract

Wi-Fi-based human activity recognition (HAR) has become an important part of integrated sensing and communications, paving the way for a range of context-aware services. However, most existing Wi-Fi-based HAR systems rely on deep learning (DL) models that are computationally and memory intensive in both training and inference, which poses significant challenges for real-world deployment. Conventional training requires simultaneous updates of millions of parameters, leading to prohibitive memory consumption. In this paper, we propose a novel quantum-assisted memory-efficient training framework (Q-MET) designed to improve efficiency in both training and inference. Q-MET utilizes a hybrid quantum classical neural network to indirectly generate parameters for HAR models, significantly reducing the trainable parameter count compared to direct optimization. To further support the deployment on resource-constrained devices, we integrate structured pruning during the training phase. Experimental results demonstrate that Q-MET achieves a 90% to 95% reduction in trainable parameters compared with conventional backpropagation-based DL training while maintaining or even exceeding classical classification accuracy. Additionally, Q-MET supports lightweight inference through structured pruning, achieving 75% to 85% model sparsity with less than 2% loss in classification accuracy. To the best of our knowledge, this work represents the first quantum-assisted approach to simultaneously tackle memory inefficiencies in both the training and inference stages of HAR systems.

Figures & tables

Symbol	Definition
$U(\theta)$	Parameterized unitary of the PQC
$U_{L}(\theta_{L})$	Parameterized unitary at the $L$ -th layer
$\|\psi(\theta)\rangle$	Output state of the PQC
$p_{\theta}(z)$	Probability of measurement outcome $z$
$z$	Possible measurement outcome
$H(n)$	Channel frequency response (CFR)

TABLE I: Summary of Notation

Tasks	Studies	Key Idea	Advantages & Limitations
HAR Accuracy Improvement	[ 56 ]	Dual-task DL with domain-independent motion features	✓Achieves high recognition accuracy in both in-domain and cross-domain settings while improving processing efficiency ✓ Supports adaptation to environmental changes without full retraining ✗ Requires complex DL architectures with high computational overhead, making deployment on resource-constrained devices challenging
	[ 57 ]	ResNet-based spatiotemporal dual-attention learning with domain-independent gesture features
	[ 58 ]	Graph-based few-shot learning with dual attention for CSI-based HAR
	[ 59 , 60 ]	Online and incremental learning for robust sensing in dynamic environments
HAR Efficiency Improvement	[ 24 ]	Channel pruning for efficient Wi-Fi CSI-based sensing	✓Reduces model complexity and parameter count while maintaining good recognition performance, enabling deployment on resource-constrained devices ✗ Primarily improves inference efficiency, with limited gains in training efficiency
	[ 32 ]	Quantization for efficient on-device Wi-Fi sensing

TABLE II: Summary of related studies with our proposed study

Fig. 1: Overview of the on-device Wi-Fi-based human activity recognition (HAR) system. (a) Dataset preparation, where CSI measurements are collected from human activities in an indoor Wi-Fi environment. (b) Training stage, where the CSI dataset is transferred to a cloud server for model training. (c) Inference stage, where the trained model is deployed on devices to perform real-time activity recognition.

Fig. 2: Architecture of the ResNet-18 model. (a) Overall architecture. (b) Initial block used in UT-HAR. (c) Initial block used in Widar3.0.

Fig. 3: Proposed quantum-assisted memory-efficient training framework. (a) Overall architecture where the Q-MET generates parameters for ResNet-18. (b) Training workflow of Q-MET: warm-up training, LAMP-based pruning, and optimal-structure training.

Fig. 4: Overview of mapping network in the proposed quantum-assisted memory-efficient training framework.

Component	Layer Type	Parameters
Inner Function †	FC (1 → 16) + BN	64
	FC (16 → 1)	17
Outer Function	FC (3 → 32) + BN	192
	FC (32 → 64)	2112
	FC (64 → 5)	325
	FC (5 → $n_{B}$ )	$6\times n_{B}$

TABLE III: Detailed architecture of mapping network model

Number of Qubits ( $N_{q}$ )	$\Delta\mathcal{C}_{\text{approx}}(\%)$
2	$-50.00$
3	$25.00$
4	$62.50$
5	$81.25$
6	$90.63$
7	$95.31$

TABLE IV: Approximate Parameter Efficiency Gains Achieved by Q-MET Across Different Qubit Counts

Datasets	UT-HAR	Widar3.0
Hardware Platform	Intel 5300 NIC	Intel 5300 NIC
Number of Classes	7	22
Sample Dimensionality	(1,250,90)	(22,20,20)
Training samples	3,977	34,926
Testing samples	996	8,726

TABLE V: Summary of Datasets

Model	Datasets	Parameters	Acc (%)
Model	Datasets	Parameters	Highest	Lowest	Average
ResNet-18	UT-HAR	11,622,599	98.80	97.60	98.08 $\pm$ 0.44
ResNet-18	Widar3.0	11,227,798	72.22	69.71	71.29 $\pm$ 1.02

TABLE VI: Baseline performance of ResNet-18 on UT-HAR and Widar3.0

QT Variants	$\Delta\mathcal{C}_{\text{approx}}(\%)$	$\Delta\mathcal{C}(\%)$	UT-HAR			Widar3.0
			Acc (%)			Acc (%)
			Highest	Lowest	Average	Highest	Lowest	Average
QT-3	25.00	24.98	99.20	99.00	99.12 $\pm$ 0.10	73.02	72.43	72.78 $\pm$ 0.22
QT-4	62.50	62.48	99.20	98.80	99.08 $\pm$ 0.16	73.14	71.82	72.35 $\pm$ 0.47
QT-5	81.25	81.22	99.40	98.80	99.20 $\pm$ 0.22	71.74	71.05	71.40 $\pm$ 0.23
QT-6	90.63	90.60	99.20	98.60	98.80 $\pm$ 0.22	72.16	69.02	70.61 $\pm$ 0.45

TABLE VII: Performance of ResNet-18 trained with QT using different numbers of qubits

Fig. 5: Training time per epoch of the baseline and QT variants.

Model	UT-HAR		Widar3.0
Model	Parameters	Acc (%)	Parameters	Acc (%)
Lightweight ResNet-18	529,295	97.60 $\pm$ 0.42	525,958	62.60 $\pm$ 0.46
Hypernetwork	547,323	98.60 $\pm$ 0.75	547,244	63.50 $\pm$ 7.39
QT-7	547,424	99.08 $\pm$ 0.24	528,830	68.76 $\pm$ 1.31

TABLE VIII: Comparison of QT-7 against classically trained lightweight ResNet-18 and a static hypernetwork

Fig. 6: Sparsity of the model on the UT-HAR and Widar3.0 datasets across varying pruning ratios.

Panel of Figure 15 — Fig. 7: Performance of Q-MET across warm-up epochs and pruning ratios on UT-HAR and Widar3.0 datasets. (a) Pruning ratio is 0.6. (b) Pruning ratio is 0.7. (c) Pruning ratio is 0.8. (d) Pruning ratio is 0.9.

Panel of Figure 16 — Fig. 8: Performance of Q-MET across varying pruning ratios on UT-HAR and Widar3.0 datasets. (a) Q-MET on UT-HAR dataset. (b) Q-MET on Widar3.0 dataset.

Panel of Figure 17 — Fig. 9: Convergence curves on the UT-HAR dataset, comparing the baseline ResNet-18 and Q-MET (7 qubits, $p=0.7$ ). The dashed line marks the pruning point at epoch 1. (a) Training loss. (b) Validation loss.

Pruning Ratio	UT-HAR				Widar3.0
Pruning Ratio	Q-MET	TTP 0	TTP 10	TTP 50	Q-MET	TTP 0	TTP 10	TTP 50
QT-7 (Unpruned)	99.08 $\pm$ 0.24				68.76 $\pm$ 1.31
0.1	99.08 $\pm$ 0.10	84.32 $\pm$ 18.42	98.36 $\pm$ 0.59	99.16 $\pm$ 0.15	68.43 $\pm$ 1.33	62.08 $\pm$ 2.73	67.65 $\pm$ 0.89	66.17 $\pm$ 1.78
0.2	98.96 $\pm$ 0.32	46.88 $\pm$ 21.82	97.84 $\pm$ 0.86	99.16 $\pm$ 0.23	69.00 $\pm$ 1.46	47.27 $\pm$ 4.77	67.46 $\pm$ 1.10	66.16 $\pm$ 1.15
0.3	99.00 $\pm$ 0.33	24.88 $\pm$ 14.44	98.00 $\pm$ 0.55	99.28 $\pm$ 0.10	68.96 $\pm$ 1.31	19.19 $\pm$ 2.89	66.97 $\pm$ 1.22	65.89 $\pm$ 1.49
0.4	98.96 $\pm$ 0.20	10.84 $\pm$ 3.86	97.04 $\pm$ 1.22	99.36 $\pm$ 0.23	68.46 $\pm$ 1.60	7.46 $\pm$ 2.94	66.34 $\pm$ 0.92	65.50 $\pm$ 1.16

TABLE IX: Comparison of Q-MET and TTP pipelines under varying pruning ratios.

Memory-Efficient Training

Memory-efficient training reduces the memory needed to optimize neural networks by compressing, recomputing, or relocating training state.

See all 90 papers

WISE-HAR: A Generalizable Ensemble Deep Learning Framework for WiFi-Based Human Activity Recognition
Jun 2, 2026Maheen Arshad, Qindeel E Zahra, Muhammad Khuram ShahzadHuman Activity Recognition Wi-Fi Sensing
Department of Computing, School of Electrical Engineering and Computer Science National University of Sciences and Technology (NUST), Islamabad, Pakistan
WiFuse: An Attention Mechanism for Human Activity Recognition using Fused CSI Amplitude and Delay-Doppler Channel Features
Aug 1, 2026Alison M. Fernandes, Hermes I. Del Monego, Bruno S. Chang +3Human Activity Recognition Multimodal Fusion
Universidade Tecnol´ogica Federal do Paran´a (UTFPR), Programa de P´os-Graduac¸˜ao em Engenharia El´etrica e Inform´atica Industrial (CPGEI-CT), Av. Sete de Setembro, 3165, Curitiba, 80230-901, Brazil · INESC TEC, Faculdade de Engenharia, Universidade do Porto, Rua Dr. Roberto Frias, Porto, 4200-465, Porto, Portugal
Wi-Spike: A Low-power WiFi Human Multi-action Recognition Model with Spiking Neural Networks
Mar 15, 2026Nengbo Zhang, Yao Ying, Lu Wang +3Human Activity Recognition Wi-Fi Sensing
School of Computing and Information Technology, Great Bay University, Dongguan, China · School of Computer Science, Guangdong University of Petrochemical Technology, Maoming, China · College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, China +2

Jun 2, 2026J/K move · Enter open · S save

WISE-HAR: A Generalizable Ensemble Deep Learning Framework for WiFi-Based Human Activity Recognition

Maheen Arshad, Qindeel E Zahra, Muhammad Khuram Shahzad

Department of Computing, School of Electrical Engineering and Computer Science National University of Sciences and Technology (NUST), Islamabad, Pakistan

Human Activity Recognition Wi-Fi Sensing

Open paper AskPDF arXiv

Human Activity Recognition (HAR) using WiFi signals has emerged as a transformative technology for smart homes, healthcare monitoring, security systems, and ambient assisted living. Unlike traditional camera-based systems that raise significant privacy concerns and fail in low-light conditions, or wearable sensors that require user compliance, WiFi-based HAR is non-intrusive, privacy-preserving, cost-effective, and works seamlessly in any lighting condition. This paper presents a comprehensive approach to recognize three distinct human activities: "No Presence" (empty room), "Walking", and "Walking + Arm-waving" using the Wallhack1.8k WiFi spectrogram dataset. We propose three key improvements to address the main challenges in WiFi-based HAR. First, to address high performance variance, we implement ensemble learning with five different CNN architectures (Deep CNN, Wide CNN, MobileNetV2, ResNet50V2, and EfficientNetB0). Second, to address the small dataset size limitation, we apply aggressive data augmentation techniques including time-warping, frequency masking, and noise addition. Third, to evaluate real-world generalization capability, we perform cross-scenario evaluation (training on Line-of-Sight and testing on Non-Line-of-Sight) and cross-antenna evaluation (training on Biquad antenna and testing on PIFA antenna). Our ensemble model achieved a test accuracy of 94.87% on the LOS scenario with Biquad antenna, outperforming the best individual model by 0.66%. Data augmentation improved Random Forest performance from 60% to 95%. Cross-scenario evaluation showed minimal accuracy drops of only 1.37% and 2.07%, demonstrating strong generalization capabilities. The results indicate that the proposed approach is robust, reliable, and suitable for real-world deployment in diverse environments with different hardware configurations.

Abstract

Memory-Efficient Training

Explore similar work

WISE-HAR: A Generalizable Ensemble Deep Learning Framework for WiFi-Based Human Activity Recognition

WiFuse: An Attention Mechanism for Human Activity Recognition using Fused CSI Amplitude and Delay-Doppler Channel Features

Wi-Spike: A Low-power WiFi Human Multi-action Recognition Model with Spiking Neural Networks