Quantum Machine Learning Benefits from Added Noise, Study Reveals

A new set of tools for quantum machine learning is emerging from National Cheng Kung University. Hsiang-Wei Huang and colleagues demonstrate the construction of both analogue and hybrid quantum kernels, revealing their ability to compete with existing kernel methods in benchmark tests and in estimating non-Markovianity from limited data. Incorporating operational noise sharply enhances kernel performance, potentially due to increased expressivity and model complexity. These findings represent a step forward in the practical application of quantum kernel methods and offer a more efficient means of assessing non-Markovianity with fewer experimental requirements.

Operational noise enhances quantum kernel performance and enables sparse data non-Markovianity

Operational noise incorporated into quantum kernels improves performance, reversing the conventional expectation that noise degrades quantum systems. This counterintuitive finding allows for estimations of non-Markovianity, a measure of a system’s dependence on its complete past, from sparser data than previously possible. Prior methods required dense datasets for accurate assessment, but these newly developed analogue and hybrid quantum kernels achieve competitive results with substantially reduced data requirements. Non-Markovianity arises in systems where the future evolution is not solely determined by the present state, but also by the history of the system; quantifying this ‘memory effect’ is crucial in diverse areas such as open quantum systems, quantum control, and quantum thermodynamics. Traditional approaches to estimating non-Markovianity often rely on calculating the Breuer-Petruccione-Komoda (BPK) measure or similar quantities, which necessitate extensive time-dependent data.

These kernels, constructed using principles of analogue quantum computing, offer a pathway towards practical quantum machine learning applications and more efficient analysis of complex systems. Analogue quantum computing leverages the natural dynamics of a quantum system to perform computations, contrasting with gate-based approaches that rely on discrete quantum gates. The analogue quantum kernel is constructed by mapping input data to the parameters of a quantum system and then measuring the overlap between the resulting quantum states. The hybrid quantum kernel combines elements of both analogue and digital computation, potentially offering the benefits of both approaches. When applied to a benchmarking dataset, the kernels achieved competitive performance against both a classical radial basis function kernel and a digital quantum kernel. Investigations showed a significant improvement over existing techniques in estimating non-Markovianity with sparse data, specifically demonstrating accurate estimations with data reduced by a factor of 5 compared to conventional methods.

The kernels’ performance characteristics were explored, revealing comparable accuracy to established methods and highlighting the potential for reduced computational cost. Incorporating operational noise improved model performance by increasing expressivity and complexity. Operational noise, in this context, refers to controlled imperfections introduced into the quantum system, such as slight variations in control pulses or environmental interactions. While seemingly detrimental, these imperfections can effectively expand the Hilbert space explored by the quantum kernel, leading to a richer representation of the input data and improved model capacity. Although current results rely on simulations and do not yet demonstrate sustained advantage on actual quantum devices with significant qubit counts, this approach represents a departure from traditional, gate-based quantum circuits. Combining elements of both analogue and digital computation offers a potential route to overcome limitations imposed by noisy intermediate-scale quantum technology, where maintaining coherence and suppressing errors is a significant challenge.

Quantum kernel methods reduce data demands for tracking system memory

Estimating how a quantum system evolves, and whether it ‘remembers’ its past, a property called non-Markovianity, has long demanded vast amounts of experimental data. This work offers a potential shortcut, demonstrating that cleverly constructed quantum kernels can extract meaningful information from far sparser datasets. The ability to accurately determine non-Markovianity with fewer data points is particularly significant for experimental quantum physics, where acquiring data can be time-consuming, expensive, and limited by the lifetime of quantum states. Dr. [Name] at [Institution] acknowledges a key gap in their work, as the study remains silent on the specific ‘other kernel methods’ used for comparison; without knowing precisely which algorithms this new approach outperforms, and by how much, it’s difficult to fully assess the significance of the advance. The benchmarking process involved comparing the performance of the quantum kernels against a standard radial basis function (RBF) kernel, a widely used classical machine learning algorithm, and a digital quantum kernel implemented on a simulated quantum computer.

Despite concerns about direct comparisons with unspecified ‘other kernel methods’, this work’s value remains substantial. A new way to analyse quantum systems has been created, potentially reducing the amount of data needed to understand how they change over time, which is important given the difficulty of gathering such data. Non-Markovianity is notoriously hard to quantify, but these quantum kernels offer a more efficient approach, potentially unlocking more practical quantum machine learning applications due to the increased model complexity achieved by incorporating operational noise. The increased complexity allows the kernel to better capture the intricate relationships within the data, leading to improved predictive power and generalisation ability. Furthermore, the reduction in data requirements could significantly accelerate the development of quantum machine learning algorithms for applications in materials science, drug discovery, and financial modelling. The researchers employed a simulation framework based on the QuTiP library to model the quantum dynamics and construct the kernels, allowing for systematic exploration of different noise levels and data sparsity conditions. The performance was evaluated using metrics such as root mean squared error (RMSE) and R-squared value, providing a quantitative assessment of the accuracy and reliability of the estimations.

Researchers successfully constructed and tested both analog and hybrid quantum kernels, demonstrating their competitive performance against existing kernel methods like the radial basis function kernel. This is significant because it offers a potentially more efficient way to estimate non-Markovianity from limited data, addressing a longstanding challenge in quantifying complex quantum system behaviour. Surprisingly, the inclusion of operational noise actually improved the kernels’ performance, likely due to increased expressivity and model complexity. The authors suggest this work provides a pathway towards practical implementations of quantum kernel methods and reduced experimental demands for estimating non-Markovianity.