Quantum vs. Classical Machine Learning: A Unified Empirical Comparison
Source: arXiv:2607.01197 · Published 2026-07-01 · By Chuanming Yu, Jiaming Liu, Zihao Ge, Xiongfei Wu, Lulu Zhu, Pengzhan Zhao et al.
TL;DR
This paper presents a first-of-its-kind unified empirical comparison between quantum machine learning (QML) models and their classical machine learning (CML) counterparts across supervised classification and reinforcement learning tasks. The authors rigorously align architectures to ensure fair comparisons, evaluating seven pairs of quantum and classical models under a shared framework using the Bars and Stripes dataset and a custom Hypercube environment. Despite initial hopes in the QML community, the results reveal that current quantum models do not outperform classical ones in overall predictive accuracy, training efficiency, or policy stability. However, quantum models show distinct advantages in specific areas such as false positive filtering (higher precision) and parameter efficiency in policy learning (fewer parameters required to achieve near-competitive performance).
Through careful experimentation under noise-free quantum simulators and constrained qubit settings, the study highlights key challenges in QML progress: limited qubit capacity necessitating dimensionality reduction, vulnerability to hardware noise, significantly slower training times due to quantum circuit overhead, and optimization instabilities such as barren plateaus. This work crystallizes the nascent state of QML in practical machine learning tasks and offers empirical evidence along with a detailed discussion of the obstacles that must be overcome to realize quantum advantage in ML. The authors provide open-source code to enable further research building on their methodology.
Key findings
- Classical SVM achieves 92.50% accuracy on Bars and Stripes vs QSVM's 82.50%, training time 0.0066s vs 11.4005s (Table 1).
- Classical CNN achieves 82.50% accuracy and 2.15s training time vs QCNN's 81.87% and 42.60s, but QCNN reaches higher precision (0.9014 vs 0.8537).
- Classical LSTM outperforms QLSTM by 8.75% accuracy and 18.5s training vs 564s training.
- In reinforcement learning, classical PG wins all episodes with average return 9.44, while QPG attains 88% win rate and return 6.94 using 4x more training time (Table 2).
- QQL uses only 66 parameters vs classical QL with 422 parameters (84.4% reduction) but achieves 94% win rate vs 100% for classical model.
- Training quantum supervised models requires roughly 10x more time; reinforcement learning training gap is even larger due to circuit execution overhead.
- Quantum models are prone to false negatives (recall lower than classical), highlighting noisier or less robust representations.
- Optimization instability and policy variance for quantum RL agents are significantly higher, indicative of barren plateaus and noisy convergence.
Threat model
The study does not focus on adversarial security but rather uses noiseless or mildly noisy simulated quantum environments as a controlled setting to evaluate QML vs classical ML under standard assumptions. There is no active adversary modeled; instead, errors and noise relate to inherent hardware imperfections and optimization difficulties.
Methodology — deep read
The paper employs a systematic experimental framework to fairly compare QML and CML across supervised classification and reinforcement learning. The threat model assumes evaluating algorithms in noiseless or limited-noise simulated quantum environments, so hardware noise and faults are studied but no active adversary is modeled.
Data provenance includes the Bars and Stripes (BAS) dataset, a canonical synthetic 4x4 binary image dataset capturing vertical/horizontal pattern classification, and a custom Hypercube Environment representing a 6-dimensional discrete state space for sequential decision-making tasks. Gaussian noise is injected into BAS data to simulate real-world variability with parameter σ=0.1. Dimensionality reduction techniques like PCA or average pooling reduce input dimensions to fit the limited 4-qubit quantum models.
Architecturally, quantum models include QSVM, QNN, QCNN, QLSTM for supervised tasks, and Quantum Policy Gradient (QPG) and Quantum Q-Learning (QQL) for reinforcement learning. Classical counterparts are carefully scaled-down networks (SVM, ANN, CNN, LSTM, PG, QL) with matched parameter counts to isolate representational differences rather than just size advantages. Quantum circuits are simulated noiselessly using PennyLane with PyTorch. Noise robustness is tested by injecting depolarizing noise in QLSTM.
Training uses Adam optimizer with fixed learning rates, batch sizes, and early stopping where applicable, running 800 episodes for RL. Parameters are tuned for fairness. Evaluation metrics across classification cover accuracy, precision, recall, F1, AUC-ROC, and training time; RL metrics include average return, win rate, learning stability (reward variance), parameter count, and training efficiency.
The authors executed all models on classical computers simulating quantum circuits—no real quantum hardware was used. Random seeds were fixed for reproducibility. Architectural alignment strictly controls parameter count and layer depth to avoid bias. Code and scripts are publicly released to enable reproducibility.
For a concrete example, in the BAS classification task, 16-pixel binary images were noise-augmented, normalized to [0, π], then PCA-reduced to 4 features for input into a 4-qubit QNN with six entangler layers. The classical ANN baseline used two hidden layers with 16 and 8 neurons. After 100 epochs of training, QNN achieved 95% accuracy but required 14.3 seconds vs ANN’s 0.66 seconds. This pipeline illustrates the overall procedure balancing rigorous comparison and practical quantum resource constraints.
Technical innovations
- A unified empirical framework for fair architectural alignment and cross-paradigm comparison of quantum and classical ML models.
- Systematic evaluation of both supervised and reinforcement learning quantum-classical model pairs under identical datasets and environments.
- Insights into parameter efficiency advantages of quantum models for RL policies, achieving competitive win rates with 80% fewer parameters.
- Identification and quantification of optimization challenges such as barren plateaus causing instability in quantum RL training.
- Detailed analysis of quantum model susceptibility to hardware noise effects simulated via depolarizing channels.
Datasets
- Bars and Stripes (BAS) — synthetic 4x4 binary images, with Gaussian noise injection σ=0.1
- Hypercube Environment — custom discrete 6-dimensional sequential decision environment with 64 states and 6 actions
Baselines vs proposed
- SVM: Accuracy = 92.50% vs QSVM = 82.50%
- CNN: Accuracy = 82.50% vs QCNN = 81.87%
- LSTM: Accuracy = 85.00% vs QLSTM = 77.50%
- Policy Gradient (PG): Average Return = 9.44 vs Quantum PG (QPG) = 6.94
- Q-Learning (QL): Win Rate = 1.00 vs Quantum Q-Learning (QQL) = 0.94
Limitations
- Use of noiseless quantum simulators instead of real quantum hardware limits assessment of practical noisy device behavior.
- Severe dimensionality reduction via PCA limits ability to test quantum models on high-dimensional, realistic datasets.
- Current qubit bottlenecks restrict quantum architectures to only 4-6 qubits, limiting expressiveness and scalability.
- Training inefficiencies and large computational overhead on classical hardware simulating quantum circuits skew practical applicability.
- Optimizations do not address advanced noise mitigation or fault-tolerant quantum computing possibilities.
- Reinforcement learning experiments limited to a synthetic, small-scale Hypercube environment rather than real-world domains.
Open questions / follow-ons
- How will quantum machine learning performance scale on fault-tolerant or error-corrected quantum hardware with more qubits?
- Can advanced noise mitigation and error correction techniques enable stable and efficient quantum model training in practical noisy devices?
- What novel quantum architectures or circuit designs can better circumvent barren plateaus and improve optimization stability?
- How do quantum models generalize to large, complex real-world datasets beyond synthetic benchmarks and low-dimensional environments?
Why it matters for bot defense
For bot-defense and CAPTCHA practitioners, this paper offers an important reality check versus hype around quantum machine learning’s immediate advantages. While QML models show potential for parameter efficiency and noise-filtering capabilities, classical techniques remain superior in accuracy, training speed, and stability under current hardware constraints. Practical bot detection systems relying on ML should remain grounded in classical approaches for now but monitor advances in quantum computing for future benefits, especially as hardware and noise resilience improve. Insights into the optimization challenges and noise sensitivity of quantum models can inform future experimental CAPTCHA defenses that consider quantum adversarial capabilities or hybrid quantum-classical detection pipelines.
Cite
@article{arxiv2607_01197,
title={ Quantum vs. Classical Machine Learning: A Unified Empirical Comparison },
author={ Chuanming Yu and Jiaming Liu and Zihao Ge and Xiongfei Wu and Lulu Zhu and Pengzhan Zhao and Jianjun Zhao },
journal={arXiv preprint arXiv:2607.01197},
year={ 2026 },
url={https://arxiv.org/abs/2607.01197}
}