Quantum continual learning on a programmable superconducting processor

Quantum computers may outperform classical computers on machine learning tasks. In recent years, a variety of quantum algorithms promising unparalleled potential to enhance, speed up, or innovate machine learning have been proposed. Yet, quantum learning systems, similar to their classical counterparts, may likewise suffer from the catastrophic forgetting problem, where training a model with new tasks would result in a dramatic performance drop for the previously learned ones. This problem is widely believed to be a crucial obstacle to achieving continual learning of multiple sequential tasks. Here, we report an experimental demonstration of quantum continual learning on a fully programmable superconducting processor. In particular, we sequentially train a quantum classifier with three tasks, two about identifying real-life images and the other on classifying quantum states, and demonstrate its catastrophic forgetting through experimentally observed rapid performance drops for prior tasks. To overcome this dilemma, we exploit the elastic weight consolidation strategy and show that the quantum classifier can incrementally learn and retain knowledge across the three distinct tasks, with an average prediction accuracy exceeding 92.3%. In addition, for sequential tasks involving quantum-engineered data, we demonstrate that the quantum classifier can achieve a better continual learning performance than a commonly used classical feedforward network with a comparable number of variational parameters. Our results establish a viable strategy for empowering quantum learning systems with desirable adaptability to multiple sequential tasks, marking an important primary experimental step towards the long-term goal of achieving quantum artificial general intelligence.

💡 Research Summary

In this work the authors present the first experimental demonstration of continual learning on a noisy intermediate‑scale quantum (NISQ) device. Using a fully programmable superconducting processor they implement a variational quantum classifier (VQC) built from 18 transmon qubits arranged in a two‑dimensional lattice. The circuit depth is 20 layers and contains 216 trainable parameters, realized with parameterized single‑qubit rotations and nearest‑neighbor controlled‑Z gates. Three sequential tasks are defined: (i) binary classification of “T‑shirt” versus “ankle‑boot” images from the Fashion‑MNIST dataset, (ii) binary classification of “hand” versus “breast” magnetic‑resonance‑imaging (MRI) scans, and (iii) discrimination of quantum many‑body states belonging to a symmetry‑protected topological (SPT) phase and an antiferromagnetic (AF) phase.

Training proceeds by minimizing a binary cross‑entropy loss for each task. Gradients are obtained experimentally via the parameter‑shift rule, which requires evaluating the circuit at shifted parameter values and directly measuring expectation values. When the VQC is trained on the tasks one after another without any regularization, the authors observe catastrophic forgetting: after learning task 2 the accuracy on task 1 collapses from ~96 % to below 30 %, and after task 3 both earlier tasks drop to near‑random performance. This confirms that, just like classical deep networks, variational quantum models are vulnerable to interference between successive data distributions.

To mitigate this, the authors adopt Elastic Weight Consolidation (EWC), a well‑known continual‑learning strategy from classical machine learning. After completing each task, they compute the Fisher information matrix for the current parameters using the same parameter‑shift measurements. The Fisher diagonal entries quantify how sensitive the loss is to each parameter, effectively measuring parameter importance. In subsequent training stages a regularization term λ ∑ₖ Fₖ(θₖ − θₖ*)² is added to the loss, penalizing changes to important parameters while still allowing plasticity for less‑critical ones. The hyper‑parameter λ is tuned for each stage.

With EWC incorporated, the quantum classifier retains high performance on all three tasks simultaneously. The average test accuracy across the three tasks exceeds 92.3 % (individual task accuracies range from 92 % to 96 %). Notably, when the same VQC is compared against a classical feed‑forward neural network that has a comparable number of trainable parameters (~200), the quantum model achieves an overall accuracy of 95.8 % on the mixed sequence of a quantum‑engineered task and a classical image task, whereas the classical network reaches only 81.3 %. The authors attribute this advantage to the high‑dimensional Hilbert space embedding provided by the quantum circuit and to the more faithful estimation of parameter importance via the quantum Fisher information.

The hardware performance is also highlighted: the processor exhibits average simultaneous single‑qubit gate fidelities of 99.96 % and two‑qubit gate fidelities of 99.68 %. These high fidelities are crucial for accurate gradient and Fisher‑information estimation, ensuring that the EWC regularization is not overwhelmed by noise. The experiment also demonstrates that the parameter‑shift rule can be efficiently applied on a real device with a manageable number of measurement shots, making the approach scalable to larger NISQ systems.

In summary, the paper shows that (1) variational quantum classifiers can be trained sequentially on distinct data domains, (2) catastrophic forgetting is a genuine issue for quantum models, and (3) Elastic Weight Consolidation, implemented with experimentally obtained Fisher information, successfully alleviates forgetting and yields a quantum advantage over a similarly sized classical model. This work establishes a practical pathway toward adaptable quantum machine‑learning systems and represents a concrete step toward the long‑term vision of quantum artificial general intelligence.