PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.04271v1cs.LG

TL;DR

DL-based Wi-Fi HAR is costly in memory during training and inference, limiting deployment on resource-constrained devices. Q-MET uses hybrid quantum-classical parameter generation and in-training structured pruning to address both stages. Across UT-HAR and Widar3.0, it reduces trainable parameters by 90% to 95% and produces 75% to 85% sparse models with limited accuracy loss.

  • Problem

    DL-based HAR requires substantial memory for parameters, optimizer states, activations, and gradients, while existing compression methods mainly target inference after full-scale training.

    Source: §I · pp. 1–2

  • Method

    Q-MET uses a hybrid quantum-classical neural network to generate classical model parameters and applies structured pruning directly during training.

    Source: §I · p. 2, A. Framework Overview · p. 8

  • Results

    Q-MET reduces trainable parameters by 90% to 95% and achieves 75% to 85% sparsity while preserving or slightly improving classification performance across UT-HAR and Widar3.0.

    Source: §I · p. 2, E. Discussion · p. 15

  • Takeaways & Limitations

    Q-MET jointly improves training and inference efficiency for compact HAR models suited to resource-constrained deployment.

    Source: §I · p. 2, §VII · p. 15

Abstract

from arXiv · show

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.

I. Introduction

Wi-Fi sensing supports practical HAR, but DL-based systems remain memory-intensive, especially during training. Q-MET addresses this gap with quantum-assisted parameter generation and pruning integrated into training for compact deployment.

  • Wi-Fi sensing extracts environmental information using existing infrastructure, supporting reliable and scalable HAR applications.Source: §I · p. 1
  • DL models learn discriminative Wi-Fi HAR representations but require substantial memory on resource-constrained edge and COTS devices.Training additionally stores parameters, optimizer states, activations, and gradients.Source: §I · pp. 1–2
  • Inference-focused compression methods typically train full-scale models before pruning or quantization, leaving training memory costs unresolved.Source: §I · p. 2
  • Q-MET uses a hybrid quantum-classical neural network to generate classical DL model parameters with fewer trainable parameters.The framework extends QuantumTrain with a quantum neural network and classical mapping network.Source: §I · p. 2
  • Structured pruning is applied during Q-MET training, producing compact models directly deployable for on-device inference without post-training compression.Source: §I · p. 2
  • Q-MET achieves up to 95% fewer trainable parameters and 75% to 85% sparsity while preserving or slightly enhancing classification performance.The evaluation uses the Widar3.0 and UT-HAR public datasets.Source: §I · p. 2

A. Wi-Fi-based Human Activity Recognition Systems

Wi-Fi HAR commonly uses CSI and DL to capture activity-related signal changes, while quantum-assisted approaches and pruning target efficiency from different directions. Q-MET-related work connects parameter-efficient training with model compression for constrained deployment.

Wi-Fi-based Human Activity Recognition Systems

Pruning

  • Prior HAR compression studies use quantization or channel pruning to reduce model size or parameters, primarily for inference efficiency.One reported approach reduced model size by up to 72%, while another reduced parameters by over 50%.Source: A. Wi-Fi-based Human Activity… · p. 3

Quantum Machine Learning

  • QML applies quantum computing properties such as superposition and entanglement to machine-learning tasks involving complex data.Source: A. Quantum Machine Learning · p. 4
  • A parameterized quantum circuit uses tunable gate parameters optimized through a hybrid quantum-classical loop.Source: A. Quantum Machine Learning · p. 4
  • A PQC produces a quantum state whose computational-basis measurement yields a probability vector for classical post-processing.For n qubits, measurement provides 2^n possible outcomes.Source: A. Quantum Machine Learning · p. 4

Pruning

  • Pruning reduces redundant parameters for deployment, but unstructured sparsity can hinder hardware acceleration because weight matrices become irregular.Source: B. Pruning · p. 4

IV. On-Device Wi-Fi–Based Human Activity Recognition Systems

The on-device Wi-Fi HAR system collects CSI from indoor activities, trains models in the cloud, and deploys compressed models for real-time recognition on resource-constrained devices.

System workflow

  • Wi-Fi-based HAR uses CSI to characterize human movements through channel fluctuations captured between wireless transmitters and receivers.CSI represents subcarrier-level amplitude and phase information derived from the channel frequency response.Source: §IV · p. 5
  • The deployment pipeline comprises dataset preparation, cloud-based training, and on-device inference.CSI is collected indoors, transferred to a cloud server for training, and used by deployed devices for real-time activity classification.Source: §IV · pp. 5–6, Figure 1 · p. 5
  • The model is compressed before deployment to reduce its size for target devices.The described compression step includes pruning before real-time inference on routers or IoT nodes.Source: §IV · pp. 5–6

Signal processing

  • The two datasets use dataset-specific input preparation before entering a shared ResNet-18 backbone.UT-HAR uses min–max normalization of 1×250×90 samples, whereas Widar3.0 uses standardization with its training-set mean and standard deviation.Source: A. Signal Processing · p. 6

B. Deep Learning Model

The study uses ResNet-18 as its representative HAR model and equips Q-MET with a quantum parameter generator that maps quantum outputs into classical network parameters.

Deep learning model

  • ResNet-18 is selected because benchmarking identifies it as highly accurate on both UT-HAR and Widar3.0.The model contains about 11.6 million trainable parameters and uses dataset-specific initial blocks for differing input shapes.Source: B. Deep Learning Model · p. 6

Quantum parameter generation

  • The Q-MET quantum parameter generator combines a parameterized quantum circuit, sinusoidal embedding, and classical mapping network.The mapping network transforms embedded quantum measurement probabilities into parameters for ResNet-18.Source: A. Framework Overview · pp. 6–8
  • The PQC uses U3 gates for qubit-state manipulation and CU3 gates for inter-qubit entanglement.Its circular layout supports multi-qubit interactions, with 6Nq trainable PQC parameters.Source: A. Framework Overview · p. 6
  • Sinusoidal embedding provides compact identifiers for computational-basis states while reducing embedding dimensionality from exponential to constant size.The embedded probabilities are represented through the triplet [S0, S1, Ψ] before mapping into classical parameters.Source: A. Framework Overview · p. 7
  • The mapping network uses inner and outer networks to produce a matrix of classical parameters.The matrix Θm has nG parameter groups with nB parameters per group and is reshaped into the ResNet-18 parameter vector.Source: A. Framework Overview · p. 8
  • Q-MET’s generated parameter vector can be reshaped for ResNet-18 and does not require CSI as input to the quantum generator.This separates parameter generation from sensing modality and supports applicability across classical DL architectures.Source: A. Framework Overview · p. 8

B. Structured Pruning

Q-MET applies structured channel pruning using layer-adaptive magnitude-based scores to remove globally low-importance channels while preserving hardware-friendly dense structures.

Structured pruning

  • LAMP-guided channel pruning derives channel importance from the LAMP scores of their associated weights.The method is intended to preserve sensitive layers more effectively than uniform or global magnitude pruning.Source: B. Structured Pruning · p. 8
  • Structured pruning is preferred because it produces regular dense matrices suited to hardware acceleration on resource-constrained platforms.Unstructured pruning instead creates irregular sparse matrices that are less practical for deployment.Source: B. Structured Pruning · p. 8

LAMP scoring

  • LAMP scores measure each weight’s relative importance against higher-magnitude weights within the same layer.The score denominator aggregates squared magnitudes of weights ranked at or above the indexed weight.Source: B. Structured Pruning · p. 8
  • Channel scores are formed by aggregating weight scores, then the lowest-scoring channels are globally removed until the target sparsity is reached.This extends LAMP from individual weights to structured channel removal across the network.Source: B. Structured Pruning · p. 8

C. Training Flow for Quantum-Assisted Memory-Efficient Training Framework

Q-MET trains a pruned model through warm-up training, structured pruning, and optimal-structure training while dynamically generating parameters for active network components.

Training flow

  • Q-MET training proceeds through warm-up training, structured pruning, and optimal-structure training.These stages establish an initial weight distribution, remove redundant filters, and continue training the compact architecture.Source: C. Training Flow for Quantum-Assisted… · p. 8
  • Warm-up training runs for a few epochs to produce a weight distribution suitable for identifying redundant connections.The warm-up stage is not intended to achieve convergence.Source: C. Training Flow for Quantum-Assisted… · p. 8

Parameter generation

2) Loss Function and Gradient Estimation:

Q-MET trains a compact quantum-classical parameter generator instead of directly updating all ResNet-18 parameters, propagating loss gradients through the generator. Its parameter efficiency improves with qubit count, exceeding 90% at six qubits while showing diminishing returns beyond that point.

Loss Function and Gradient Estimation

  • Q-MET propagates the loss gradient through the parameterized quantum circuit and mapping network that generate ResNet-18 parameters.The gradient uses the Jacobian of generated ResNet-18 parameters with respect to the joint generator parameters.Source: 2) Loss Function and Gradient Estimation · p. 9
  • Q-MET restricts trainable updates to the compact quantum parameter generator, avoiding direct training of the massive ResNet-18 parameter set.This design reduces training memory overhead in resource-constrained environments.Source: D. Parameter Efficiency Analysis · p. 9

Parameter Efficiency

  • 90.63% parameter efficiency is achieved with Nq = 6, while Nq = 2 produces a −50% gain and Nq ≥3 reaches approximately 25%.The six-qubit result is derived from the reported relationship between qubit count and parameter efficiency.Source: D. Parameter Efficiency Analysis · p. 9
  • 99.41% maximum parameter efficiency is reached at Nq = 10, but increasing Nq from 6 to 10 adds only 8.78% gain.The reported trend indicates diminishing returns after six qubits.Source: D. Parameter Efficiency Analysis · p. 10

E. Computational Complexity Analysis

Q-MET combines quantum parameter generation, mapping-network computation, and pruned ResNet-18 execution, trading additional computation for substantially lower trainable-state memory. The experiments use two public CSI-based HAR datasets spanning human motions and fine-grained gestures.

Complexity Components

  • Q-MET’s per-iteration complexity combines PQC cost O(2^Nq), mapping-network cost O(nGnB), and pruned-model cost O((1−s)FResNet-18).The sparsity-dependent term scales with the remaining fraction of the full ResNet-18 training cost.Source: E. Computational Complexity Analysis · p. 10

Training-Time Trade-off

  • 18%–36% higher training time is observed because Q-MET adds parameter-generation overhead relative to standard full-model backpropagation.Standard training has complexity O(FResNet-18).Source: E. Computational Complexity Analysis · p. 10

Memory Efficiency

  • Q-MET’s main benefit is memory efficiency: trainable-state memory scales with O(CQ-MET), where CQ-MET ≪ CResNet-18.Standard training stores weights, gradients, and optimizer states for all ResNet-18 parameters.Source: E. Computational Complexity Analysis · p. 10

Evaluation Datasets

  • The evaluation uses UT-HAR and Widar3.0, both containing CSI collected with an Intel 5300 Wi-Fi interface under real-world conditions.UT-HAR covers basic human motions, whereas Widar3.0 targets fine-grained gesture recognition.Source: A. Dataset · p. 10

B. Training Hyperparameter Configuration

The study defines accuracy, accuracy loss, parameter count, training memory, sparsity, and average epoch time as evaluation measures. Q-MET reduces model-state training memory substantially, while activation and buffer memory remain outside the framework’s primary target.

Performance Metrics

  • Accuracy loss compares Q-MET and baseline ResNet-18 through AccQ-MET and Accbaseline.Accuracy itself is defined as the ratio of correct predictions to total test instances.Source: 1) Accuracy · p. 10

Efficiency Metrics

  • Trainable-parameter count is used as the primary efficiency metric for comparing Q-MET with classical training.The paper identifies parameter count as a canonical measure in parameter-efficient training.Source: 2) Parameter Efficiency · pp. 10–11

Training Memory

  • 95% and 90% model-state memory reductions are reported for UT-HAR and Widar3.0, respectively.Baseline memory is approximately 177 MB versus 8.4 MB for QT-7 on UT-HAR, and 171 MB versus 17 MB for QT-6 on Widar3.0.Source: 2) Parameter Efficiency · p. 11

Memory Scope

  • Activation and buffer memory is orthogonal to the proposed framework’s target, although structured pruning incidentally reduces it.The reported memory figures quantify parameters, gradients, and optimizer states under FP32.Source: 2) Parameter Efficiency · p. 11

Additional Metrics

  • Sparsity measures the percentage of the model pruned, while average training time per epoch is measured until convergence.Higher sparsity generally corresponds to reduced model size and faster inference.Source: 3) Sparsity · p. 11

3) Sparsity:

The section frames sparsity as a pruning metric and training time as an efficiency measure, while describing experiments across training configurations. The reported setup compares baseline ResNet-18 performance on UT-HAR and Widar3.0.

Sparsity

  • Higher sparsity indicates more aggressive structured pruning and typically corresponds to reduced model size and faster inference.Sparsity is defined as the percentage of the model that has been pruned.Source: 3) Sparsity · p. 11

Training Time

  • Average training time per epoch is defined as total training time divided by the number of epochs until convergence.This metric is used to evaluate QT computational efficiency.Source: 4) Training Time · p. 11

Experimental Settings

  • The experiments vary Warm-Up epochs and pruning ratios to analyze Q-MET performance-efficiency trade-offs.The evaluation also examines QT configurations before studying Q-MET settings.Source: D. Experimental Results · p. 11

Baseline Performance

  • Table VI reports baseline ResNet-18 performance on UT-HAR and Widar3.0.The table is identified as a baseline-performance summary for the two datasets.Source: Table VI · p. 11

1) Performance of Classical ResNet-18 Model and QT Framework:

QT reduces ResNet-18 trainable parameters substantially while preserving or improving UT-HAR accuracy, though aggressive compression degrades performance on complex Widar3.0 data. The framework also incurs higher training time, while pruning produces high sparsity.

Classical baseline

QT performance

  • 95.29% parameter reduction with QT-7 accompanied 99.08 ± 0.24% average UT-HAR accuracy, exceeding the baseline.QT-5 also reduced parameters by approximately 81.22% while reaching 99.20 ± 0.22%.Source: 1) Performance of Classical ResNet-18… · p. 11

Training efficiency

  • 35.7% higher training time at QT-10 than the baseline accompanied the parameter-efficiency gains.The baseline required 3.2 seconds per epoch, while QT-3 already incurred an 18.2% overhead.Source: 1) Performance of Classical ResNet-18… · p. 12

QT performance

  • 90%+ parameter reduction with QT-6 preserved 70.61 ± 0.45% accuracy on Widar3.0, whereas QT-9 and QT-10 fell to 66.28 ± 1.94% and 65.84 ± 2.03%.Increasing variance at higher qubit counts indicates reduced stability and possible underfitting.Source: 1) Performance of Classical ResNet-18… · p. 12

Controlled comparisons

  • QT-7 outperformed a parameter-matched lightweight ResNet-18 by 1.48 percentage points on UT-HAR and 6.16 points on Widar3.0.The comparison used a slim ResNet-18 retaining 22% of each convolutional layer’s channels.Source: 2) Comparison with Lightweight ResNet-18 · p. 12

Pruning

  • 90% sparsity was reached at a pruning ratio of 0.9, with sparsity exceeding 70% above a ratio of 0.6.Sparsity increased monotonically with the pruning ratio on both datasets.Source: 4) Impact of Pruning Ratios and… · p. 13

5) Q-MET Performance:

Q-MET integrates structured pruning into quantum-assisted training to improve deployment efficiency while retaining competitive recognition accuracy. Across datasets, it achieves high sparsity with modest accuracy losses, although Widar3.0 is more sensitive to aggressive pruning.

Pruning performance

  • 98.44%+ UT-HAR accuracy was maintained at a 0.6 pruning ratio while more than 70% of parameters were removed.Accuracy remained above 98.4% across pruning ratios from 0.1 to 0.8.Source: 5) Q-MET Performance · p. 13
  • 70.02% peak Q-MET accuracy on Widar3.0 at a 0.2 pruning ratio was close to the QT-6 baseline’s 70.61%.Accuracy declined to 66.01% at a pruning ratio of 0.9, with higher variance under extreme sparsity.Source: 5) Q-MET Performance · p. 13

Train-then-prune comparison

  • 50 extra fine-tuning epochs were required for TTP50 to approach Q-MET at a 0.7 pruning ratio, while Q-MET needed no post-pruning training.At that ratio, TTP50 reached 98.80% on UT-HAR and 64.74% on Widar3.0, compared with Q-MET’s 98.84% and 66.72%.Source: 6) Comparison with Train-Then-Prune… · p. 14

Inference efficiency

  • Q-MET reached 85%+ UT-HAR sparsity at a 0.8 pruning ratio with 98.60% accuracy, within 1% of unpruned QT-7.The pruned model still exceeded the classical unpruned baseline’s 98.08%.Source: E. Discussion · p. 15
  • 75% Widar3.0 sparsity at a 0.6 pruning ratio yielded 69.20% accuracy, 1.42 points below unpruned QT-6.A single warm-up epoch was sufficient for steady pruning performance.Source: E. Discussion · p. 15

Overall results

  • 95% fewer trainable parameters on UT-HAR and 90% fewer on Widar3.0 were reported alongside approximately 85% and 75% model sparsity, respectively.Accuracy improved from 98.08% to 99.08% on UT-HAR and decreased from 71.29% to 70.61% on Widar3.0.Source: §VII · p. 15, E. Discussion · p. 15

Related papers

Loading 2609.04271v1…