🤖 machine learning

Quantum vs. Classical Machine Learning: A Unified Empirical Comparison

This paper presents a unified empirical comparison of seven quantum and classical machine learning model pairs, finding that while current quantum models do not yet outperform classical baselines in prediction, stability, or training time, they show promise for noise filtering and false positive control, while highlighting key challenges in hardware, efficiency, and convergence.

Original authors: Chuanming Yu, Jiaming Liu, Zihao Ge, Xiongfei Wu, Lulu Zhu, Pengzhan Zhao, Jianjun Zhao

Published 2026-07-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Chuanming Yu, Jiaming Liu, Zihao Ge, Xiongfei Wu, Lulu Zhu, Pengzhan Zhao, Jianjun Zhao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine two racing teams: Team Classical (the veterans who have been running marathons for decades) and Team Quantum (the new, high-tech runners using futuristic, gravity-defying shoes).

This paper is a head-to-head race report. The authors set up a fair track where both teams run the exact same distance, with the same rules, to see who actually wins right now.

Here is the breakdown of their findings in plain English:

The Setup: A Fair Race

To make sure the comparison was honest, the researchers didn't let the "Classical" team use a massive truck while the "Quantum" team used a bicycle. They forced both teams to use vehicles of similar size and complexity.

The Track: They used two types of courses. One was a pattern-matching game (identifying bars and stripes), and the other was a maze navigation game (finding a way through a 64-step cube).
The Rules: They ran the race on a perfect, noise-free computer simulation (like a video game) to see what the technology could do theoretically, without the messy interference of real-world broken hardware.

The Results: Who Won?

1. The Overall Scoreboard: Team Classical Wins
In almost every category, the classical team finished faster and more accurately.

Speed: The classical models trained (learned) dozens of times faster. Some quantum models took hundreds of times longer to finish the same task.
Accuracy: When it came to getting the right answer, the classical models were more consistent. They made fewer mistakes overall.
Stability: The classical runners were steady. The quantum runners sometimes stumbled, got confused, or took a wrong turn during training, making their performance unpredictable.

2. The Silver Lining: Where Team Quantum Shined
Even though they lost the race, the Quantum team showed some unique superpowers in specific situations:

The "Noise Filter" Superpower: In the pattern-matching game, the Quantum team was surprisingly good at ignoring "false alarms." If the data was a bit messy, the Quantum models were better at saying, "No, that's not a match," whereas the classical models sometimes guessed wrong. They were more careful.
The "Efficient Backpack" Superpower: In the maze game, the Quantum team achieved nearly the same success rate as the classical team but with a tiny backpack. The classical models needed hundreds of "parameters" (like memory slots or brain cells) to solve the maze. The Quantum models solved it with only a fraction of that. They packed a lot of power into a very small space.

The Hurdles: Why Quantum Isn't Winning Yet

The paper explains that Team Quantum is currently held back by four major problems:

The "Tiny Room" Problem: Quantum computers currently have very few "qubits" (the basic units of information). It's like trying to fit a whole library into a shoebox. To make the data fit, researchers have to crush it down (compress it), which risks losing important details.
The "Fragile Glass" Problem: Quantum systems are incredibly sensitive. A tiny bit of noise or interference (like a slight vibration) can break the calculation. They are like glass sculptures in a windstorm.
The "Slow Motion" Problem: Because quantum calculations are so complex to simulate, they take forever to run. It's like trying to paint a masterpiece by moving one brushstroke per hour, while the classical team paints by the minute.
The "Flatland" Problem: When the quantum models try to learn, they often get stuck in a "barren plateau." Imagine trying to climb a mountain, but the ground is perfectly flat for miles. The model doesn't know which way is "up," so it wanders aimlessly instead of improving.

The Bottom Line

The paper concludes that Quantum Machine Learning is not yet ready to beat the classics. If you need a model that is fast, accurate, and stable today, stick with classical methods.

However, Quantum learning is a promising new tool. It has a special talent for filtering out noise and doing a lot with very little memory. The researchers believe that once we fix the hardware (make the "shoes" more durable) and the algorithms (teach the runners how to climb the flat mountains), Quantum might eventually take the lead. But for now, it's still a work in progress.

Technical Summary: Quantum vs. Classical Machine Learning: A Unified Empirical Comparison

Problem Statement

Despite the theoretical potential of quantum computing to offer computational advantages in machine learning (ML), the current empirical evidence supporting the superiority of Quantum Machine Learning (QML) models over classical counterparts remains insufficient. While numerous QML algorithms and frameworks have been proposed, it is unclear whether they offer genuine advantages in overall generalization, training efficiency, or stability compared to traditional models. Existing surveys often lack a unified, side-by-side empirical comparison that controls for architectural variables, leaving a gap in understanding the practical applicability and limitations of QML in the Noisy Intermediate-Scale Quantum (NISQ) era.

Methodology

To address this gap, the authors conducted a systematic empirical study comparing QML and Classical Machine Learning (CML) models under a unified framework. The study utilized the PennyLane framework integrated with PyTorch and Scikit-learn, executing all experiments on a noiseless state-vector quantum simulator to isolate representational power from hardware noise, while strictly aligning hyperparameters and network architectures to ensure fair comparison.

Experimental Design

The study evaluated seven model pairs across two primary domains:

Supervised Learning (Classification):
- Dataset: A "Bars and Stripes" (BAS) dataset, chosen for its structured patterns suitable for quantum encoding.
- Models: Five pairs were compared: QSVM vs. SVM, PegasosQSVM vs. PegasosSVM, QNN vs. ANN, QCNN vs. CNN, and QLSTM vs. LSTM.
- Constraints: To address qubit limitations, classical models were deliberately restricted in capacity (e.g., reduced hidden layers, limited neurons) to match the parameter scale of their quantum counterparts. Quantum inputs were normalized to $[0, \pi]$ and dimensionality was reduced via PCA or pooling to fit 4-qubit constraints.
Reinforcement Learning (Sequential Decision-Making):
- Environment: A custom "Hypercube Environment" with 64 discrete states and 6 actions, designed to map directly to a 6-qubit quantum state space via angle embedding.
- Models: Two pairs were compared: Quantum Policy Gradient (QPG) vs. Classical PG, and Quantum Q-Learning (QQL) vs. Classical Q-Learning (QL).
- Constraints: Classical baselines were implemented as neural networks with parameter counts and optimization strategies (e.g., $\epsilon$ -greedy decay, replay buffers) strictly aligned with the quantum agents.

Evaluation Metrics

The study employed a comprehensive set of metrics:

Supervised: Accuracy, Precision, Recall, F1-Score, AUC-ROC, and Training Efficiency (time).
Reinforcement Learning: Average Return, Win Rate, Learning Stability (variance of rewards), Parameter Count, and Training Efficiency.

Key Contributions

Unified Empirical Comparison: This work provides the first systematic comparison of QML and CML covering seven model pairs across both supervised and reinforcement learning tasks under a single, unified framework.
Architectural Alignment: The study introduces a rigorous evaluation methodology that isolates the true representational power of quantum models by restricting classical baselines to comparable parameter scales, avoiding the bias of massive classical parameter advantages.
Identification of Specific Advantages and Challenges: The paper delineates specific conditions where QML shows promise (e.g., noise filtering) and catalogs the critical barriers hindering its advancement.

Empirical Results

1. Overall Performance (RQ1)

The results indicate that classical models currently outperform QML models in overall predictive accuracy, policy stability, and training efficiency.

Classification: Classical baselines achieved higher accuracy, F1-scores, and AUC-ROC across all pairs. For instance, the classical SVM (0.9250 accuracy) outperformed QSVM (0.8250), and the classical LSTM (0.8500) surpassed QLSTM (0.7750).
Reinforcement Learning: Classical agents consistently achieved perfect win rates (1.0000) and higher average returns compared to their quantum counterparts (QPG: 0.8800 win rate; QQL: 0.9400 win rate).
Training Efficiency: A fundamental disparity exists in computational costs. Classical models trained significantly faster. The training time for quantum models was orders of magnitude higher (e.g., QLSTM took ~564 seconds vs. ~18 seconds for LSTM; QPG took ~2150 seconds vs. ~6 seconds for PG).

2. Specific Advantages of QML (RQ2)

Despite trailing in overall metrics, QML demonstrated distinct advantages under specific configurations:

Precision and False Positive Control: The Quantum Convolutional Neural Network (QCNN) achieved notably higher precision (0.9014) compared to the classical CNN (0.8537), suggesting quantum encoding mechanisms may be more effective at filtering noise and controlling false positives in spatial data.
Parameter Efficiency: In reinforcement learning, quantum agents achieved competitive performance with significantly fewer parameters. The QQL agent reached a 94% win rate with only 66 parameters, an 84.4% reduction compared to the classical baseline (422 parameters). This suggests high representational density in parameterized quantum circuits.

3. Challenges (RQ3)

The study identifies four critical challenges hindering QML advancement:

Qubit Bottlenecks: The necessity for strict dimensionality reduction (e.g., PCA, pooling) to fit data into limited qubit counts restricts the ability to leverage high-dimensional raw data, potentially causing information loss.
Hardware Noise and Robustness: QML models are highly sensitive to environmental imperfections. While evaluated on simulators, the study notes that transitioning to physical hardware requires robust noise mitigation, as models are prone to performance degradation under noise.
Training Efficiency Gap: The computational overhead of QML is severe, particularly in reinforcement learning where every step requires circuit execution, state preparation, and measurement. This creates a massive bottleneck compared to classical vector operations.
Optimization Stability: Quantum agents exhibited high variance and instability during policy convergence, likely due to the "Barren Plateaus" phenomenon and stochastic noise, making them less reliable for stable training than classical algorithms.

Significance and Claims

The paper claims that while QML models do not yet surpass classical baselines in overall generalization, training speed, or stability, they remain a promising approach for specific tasks such as filtering noise and controlling false positives. The research provides a foundational summary of the challenges facing the field, specifically regarding hardware constraints, noise interference, and convergence stability.

The authors conclude that future research must focus on designing noise-resilient quantum architectures and developing robust parameter-optimization protocols to overcome current limitations. The work serves as a realistic benchmark, clarifying that while QML holds theoretical potential, its practical superiority is currently limited by the constraints of the NISQ era and algorithmic instability.