Quantum neural networks may need fewer cloud jobs with new method

Muhammad Faryad of Lahore University of Management Sciences (LUMS) and Rabab Alomairy of King Abdullah University of Science and Technology (KAUST) demonstrate a new approach to training quantum neural networks that can significantly reduce the number of interactions with a quantum processor. Their work reframes hybrid quantum-classical training as an execution-efficiency problem, focusing on maximizing optimization progress with each costly job submitted to a cloud queue.

Simulations reveal that a batched parameter-shift training method requires only a fifth to half the number of jobs needed by simultaneous-perturbation methods (SPSA) for models with 24 to 48 parameters. The researchers found that the gradient error of their batched method falls as the number of shots per job increases, unlike SPSA which has a persistent error floor.

Batched Parameter-Shift Training Reduces Quantum Neural Network Jobs

This efficiency stems from a focus on execution, treating hybrid quantum-classical training as a problem of maximizing progress from each interaction with a quantum processor. This contrasts sharply with simultaneous-perturbation methods, known as SPSA, which exhibit a gradient error floor that cannot be overcome by simply increasing the number of shots taken per calculation. In simulated environments, the batched method consistently outperformed SPSA across a parameter range of 24 to 48, maintaining this advantage even when accounting for two levels of noise.

On IBM’s ibm_fez processor, the batched optimizer reached SPSA’s final training loss within four to 20 jobs, in three interleaved runs using six and eight qubits. The efficiency gains extend beyond job count, impacting overall time spent accessing quantum resources. This is particularly relevant given that each job submitted to a cloud queue incurs both a wait time and inherent overhead before execution begins.

The researchers note the importance of maximizing the information extracted from each costly quantum interaction. The structure-aware stencil search, a variation of the batched method, also outperformed SPSA, further demonstrating the potential of this approach to optimize quantum neural network training.

Optimizer Comparison: Jobs vs. Circuit Evaluations for QNNs

This optimization focuses on how to best use each job submission, rather than simply minimizing the number of circuit evaluations or shots per job, framing the problem as an execution-efficiency challenge. The approach acknowledges that each round trip to a quantum processor incurs both a wait time and a cost, making the effective use of each submission paramount.

This work builds on a batched pattern search initially developed for the quantum approximate optimization algorithm (QAOA), but notes that a direct application of that method stalls when scaling to quantum neural networks with more than a few dozen parameters. The researchers found that the number of jobs needed to achieve a target solution quality is the critical metric, as it directly translates to time and resource expenditure.

The researchers write, emphasizing the importance of grouping circuit evaluations strategically. This approach mirrors techniques used in high-performance computing to amortize communication and launch overheads, adapting those principles to the unique constraints of cloud-based quantum processing. The study’s findings suggest that optimizing job allocation can significantly reduce the overall cost and time required for training quantum neural networks.

Parameter-Shift Stencil vs. SPSA Gradient Error Analysis

The error rate of batched parameter-shift training decreases proportionally to the number of shots per job, a quantifiable improvement over alternative optimization methods, according to new simulations and hardware tests. Analysis reveals that SPSA spends most of each step calculating the gradient, incurring a fixed error proportional to the number of trainable angles and the gradient itself. This approach sends current parameter values alongside shifted copies for each angle, enabling parallel processing of all 2P + 1 evaluations without dependency between rows.

While barren plateaus, regions where gradients vanish, limit all methods equally, job-awareness significantly impacts how quickly a solution can be found within a reachable region. The study emphasizes that the classical optimizer’s decision on how much learning occurs per round trip is critical when using a queued quantum processor, as the cost is determined by each interaction with the quantum hardware.

Structure-Aware Sinusoid Search Improves Optimization Efficiency

A new optimization technique, sinusoid stencil search, improves efficiency by treating each job sent to a quantum processor as a batch, rather than a series of individual calculations. This performance was observed across three interleaved runs, suggesting consistent results despite the inherent noise in quantum hardware.

As the source explains, “Because each angle enters once as a Pauli rotation, the output of every data point is a sinusoid in that angle,” allowing the search to efficiently locate the minimum loss value. The study’s results hold consistent across two noise levels, reinforcing the robustness of the method. Its test accuracy was clearly higher at eight qubits; at six qubits the difference was not significant.

Job Model Defines Cost in Hybrid Quantum-Classical Training

The efficiency of training quantum neural networks is directly tied to how jobs are structured and submitted to quantum processors, with one new method requiring significantly fewer interactions than standard approaches. This reduction in job count translates to substantial savings in metered time, the period the quantum processor is exclusively allocated, and overall round-trip time, a critical factor when using cloud-based quantum resources.

Analyzing the cost of each job reveals a breakdown of waiting time, metered time and the time required for circuit evaluations. Researchers measured the metered time of 306 jobs, finding an initial cost of approximately 2.0 seconds, with additional time accrued at a rate of 0.27 milliseconds per shot and 0.5 milliseconds per row of data.

The study focused on a data re-uploading circuit built natively on a chain of physical qubits, using a classifier task designed to assess training speed rather than overall model performance against classical algorithms. This controlled workload allowed for a direct comparison of optimization strategies, isolating the impact of job structure on training cost. All hardware numbers can be re-derived from the raw shots.

The researchers emphasize that evaluating optimizers based on quality, job count and metered cost, rather than solely on the number of shots, provides a more practical metric for assessing hybrid training efficiency. This approach highlights how strategic job design can significantly reduce the financial and temporal burden of using quantum processors for machine learning tasks. This difference in job count translates to a more efficient use of quantum processing resources.

Pauli Rotations Enable Efficient Job-Aware Parameter Batching

A batched parameter-shift method uses the structure of circuits built from Pauli rotations to improve training efficiency, particularly as model complexity increases. The core principle is to pre-define a batch of vectors before submission, a technique that transfers from successful implementations in quantum approximate optimization algorithms. This new optimizer, termed a sinusoid stencil search, exploits the sinusoidal nature of outputs from Pauli rotations. Each angle within a circuit generates a sinusoidal response, and the method reads three rows of parameter shifts to define a complete sinusoid.

By probing every coordinate simultaneously, the search minimizes additional round trips to the quantum processor, reducing overhead. The method constructs each job, denoted as $Q_k$, as a list of parameter vectors paired with the minibatches they are evaluated on, with the job-aware aspect being that $Q_k$ depends solely on the results of prior jobs. The efficiency gains stem from how parameter vectors are grouped.

Each job contains two points, θ ± $c_k{\Delta }_k$, where ${\Delta }_k$ is a random vector of ± 1 entries, evaluated on minibatches of 64 points. The resulting gradient estimate, g, then informs an update to the parameter values θ. “The job-aware part is that all 2P + 1 parameter vectors are known before the job is sent, so they travel together,” explains the research.

More like this

Artificial IntelligenceMultiverse Computing tests quantum neural networks in a 30B LLM

High Performance ComputingPower grid problem tests Quantum Inspire’s link to DelftBlue

Quantum AlgorithmsSymmetry in quantum neural networks boosts accuracy on image tasks

Quantum Machine LearningResearchers map qubit states to probability distributions in phase space