Image Classification on IBM Quantum Computers

Research Question and Approach

This paper addresses the practical challenges of deploying multi-class machine learning models on Noisy Intermediate-Scale Quantum (NISQ) hardware. While quantum machine learning (QML) is theoretically promising, real-world execution is hindered by high training costs, inefficient hardware utilization during inference, and the difficulty of scaling beyond binary classification. The authors propose a unified framework to classify ten-class MNIST digits on IBM’s 127-qubit Eagle processor. Their approach centers on a two-phase training protocol: first, a classical encoder and readout are trained via gradient descent on a noiseless simulator; second, the quantum circuit parameters are optimized using a gradient-free algorithm (COBYLA) to avoid the prohibitive costs of parameter-shift gradients on hardware.

Key Contributions

The study introduces three major improvements to the QML pipeline. First, it decouples classical and quantum optimization, demonstrating that on-hardware fine-tuning yields no significant accuracy improvement over training on a simulator. This finding shifts the focus toward an inference-only deployment strategy. Second, the authors implement Quantum Multi-Programming (QMP), which packs multiple circuit copies onto a single device. This technique significantly increases throughput by reducing the number of quantum job submissions, which are typically dominated by queue latency rather than computation time. Third, the authors perform a controlled comparison of five different circuit architectures (ansätze), providing a benchmark for practitioners to balance classification accuracy against compilation costs like gate depth and CNOT counts.

By systematically addressing the bottlenecks of training, inference, and architecture selection, this work provides a blueprint for practical QML deployment on current hardware. The demonstration that training can be offloaded to classical simulators while reserving quantum hardware for inference provides a clear, scalable workflow for researchers. While the authors note that their quantum model does not currently outperform a classical network of similar capacity, the framework establishes a necessary foundation for future experiments aiming to achieve genuine quantum advantage in multi-class image classification.