Noisy Data Still Allows Quantum Machine Learning to Function

Imperfections in both classical data and quantum hardware impact the performance of quantum machine learning models. Bhavna Bose and colleagues at SVKM’s NMIMS Mukesh Patel School of Technology Management and Muhammad Faryad from Lahore University of Management show that noise from flawed classical data can worsen the effects of quantum decoherence. Their thorough study, utilising the Titanic dataset and a variety of noise models within the Qiskit Aer simulator, reveals a sharp reduction in classification accuracy when both classical and quantum noise are present. These findings highlight the key need to account for combined classical and quantum noise when developing and evaluating quantum machine learning pipelines for near-term applications.

Simulating data corruption and quantum noise for strong machine learning

A carefully constructed methodology for simulating realistic noise conditions within a quantum machine learning pipeline formed the basis of this work. Classical data initially underwent corruption via several models, including speckle, impulse, quantization, and feature dropout, representing common real-world imperfections encountered during data acquisition and pre-processing. Speckle noise, mimicking sensor inaccuracies, introduces multiplicative noise, while impulse noise simulates sudden, transient errors like bit flips. Quantization represents the loss of precision when converting continuous data into discrete values, and feature dropout randomly removes input features, simulating missing data scenarios. This flawed data then passed through a ZZ feature map, a technique for converting classical information into a quantum format understandable by the quantum computer. The ZZ feature map encodes classical data into quantum states by applying rotations around the Z-axis of qubits, effectively creating a superposition of states representing the input features. This is an important step because any initial errors are translated into the quantum realm, potentially amplifying their impact during subsequent quantum computations.

The Titanic dataset, a commonly used benchmark in machine learning, was employed to benchmark a variational quantum classifier against multiple noise types. This dataset contains information about passengers aboard the Titanic, used to predict survival based on features like age, sex, and class. After the ZZ feature map conversion, the data was subjected to quantum noise channels including depolarizing, amplitude, phase damping, Pauli, and readout errors. Depolarizing noise randomly applies an identity operation to a qubit, effectively erasing its quantum state. Amplitude and phase damping represent energy loss from the qubit, leading to decoherence. Pauli errors introduce bit-flip or phase-flip errors, and readout errors occur during the measurement of the qubit’s state. Qiskit Aer simulator was utilised to model these quantum effects, allowing for a detailed examination of how these errors propagate through the system and affect classification outcomes. The simulator allows researchers to control the level of each noise type, enabling a systematic investigation of their combined impact. By simulating these errors, the researchers could assess the robustness of the variational quantum classifier without requiring access to actual noisy quantum hardware.

Combined classical and quantum noise severely impairs variational quantum classification

Classification accuracy, utilising the Titanic dataset, dropped to 78.5% when both classical and quantum noise were applied, a reduction of over 15 percentage points compared to simulations with no noise. This threshold represents a key point where the variational quantum classifier’s ability to reliably distinguish between classes becomes severely compromised. Previously, achieving accurate classification under such combined noise conditions was considered impossible. Classical input noise exacerbates quantum decoherence, creating unstable training dynamics and sharply diminishing the potential benefits of near-term quantum machine learning. Variational quantum classifiers rely on iterative optimisation of quantum circuit parameters, and noise can disrupt this process, leading to suboptimal solutions and reduced accuracy.

Modelling realistic noise at multiple levels, from data acquisition to quantum circuit execution, provides a more pragmatic assessment of QML performance than previous studies. Applying four distinct classical noise types, speckle, impulse, quantization, and feature dropout, to the Titanic dataset before quantum processing revealed that even moderate classical corruption significantly impacted performance. Specifically, a 10% level of impulse noise reduced classification accuracy to 72.3%, demonstrating a clear sensitivity to input data quality beyond quantum errors. This highlights that the quality of the initial data is paramount, even before considering the limitations of quantum hardware. Simulations incorporating both Pauli errors and amplitude damping at the quantum circuit level, alongside speckle noise in the classical data, resulted in a combined accuracy of just 69.8%. This highlights a combination of degradation not observed when noise sources were tested in isolation. The synergistic effect of combined noise suggests that mitigating one type of noise is insufficient; a holistic approach is required. While these results demonstrate a substantial performance drop under realistic conditions, scaling these findings to larger, more complex problems and diverse hardware platforms remains a vital next step. Investigating the impact of these noise combinations on datasets with higher dimensionality and more intricate feature relationships will be crucial for assessing the generalizability of these findings.

Classical data imperfections limit variational quantum classifier performance

Establishing reliable quantum machine learning demands more than simply refining the quantum hardware itself. Researchers at SVKM’s NMIMS Mukesh Patel School of Technology Management, led by Bhavna Bose, convincingly demonstrate that even pristine qubits are vulnerable to errors originating in the classical data used to train the algorithms; a variational quantum classifier, a core component of this emerging field, is particularly sensitive. The variational quantum classifier is a hybrid quantum-classical algorithm that uses a quantum circuit to prepare a quantum state and a classical optimizer to adjust the circuit parameters to minimise a cost function. Dr. Bose acknowledges that their analysis relied on a single, well-known dataset, the Titanic passenger list, raising questions about whether these findings hold true for more complex real-world problems. Further research is needed to evaluate the robustness of these findings across a wider range of datasets and problem domains.

Imperfections in classical data exacerbate quantum decoherence within variational quantum classifiers, impacting training stability and reducing classification accuracy. By systematically combining realistic classical noise with simulated quantum hardware errors, a synergistic degradation of performance was demonstrated, highlighting the importance of data quality alongside hardware improvements for reliable machine learning. This work underscores that addressing classical data imperfections is crucial for realising the full potential of quantum machine learning algorithms. Further investigation into robust data pre-processing techniques, such as error correction codes and noise filtering algorithms, and noise mitigation strategies will be essential for building practical and reliable QML systems. Exploring the use of data augmentation techniques to increase the robustness of the model to noisy data is also a promising avenue for future research. Ultimately, a comprehensive approach that addresses both classical and quantum noise sources will be necessary to unlock the full potential of quantum machine learning.

The research demonstrated that imperfections in classical data can worsen the effects of quantum decoherence in variational quantum classifiers. This matters because it indicates that the quality of input data is as important as the quantum hardware itself for achieving reliable machine learning results. Using the Titanic dataset, researchers showed that combining classical noise with simulated quantum errors led to decreased classification accuracy and unstable training. The authors suggest future work should focus on improving data pre-processing and noise mitigation strategies to build more robust quantum machine learning systems.