Quantum continual learning is demonstrated on superconducting qubits.

A superconducting-qubit processor was used to demonstrate quantum continual learning on three sequential tasks, and applying elastic weight consolidation helped the quantum classifier retain knowledge with an average accuracy reported above 92.3%.

A research team reports an experimental demonstration of quantum continual learning on a superconducting qubit processor. The manuscript is provided in an unedited pre-publication form and may be revised before final publication. The authors sequentially trained a quantum classifier on three tasks — two involving real-world images and one involving quantum states — and observed catastrophic forgetting during training without mitigation. They then applied an elastic weight consolidation strategy to reduce forgetting and retain prior knowledge. Key findings: - The experiment used a superconducting-qubit processor and sequentially trained a quantum classifier on three tasks (two image tasks, one quantum-state task). - Catastrophic forgetting was observed when the model was trained sequentially without mitigation. - Applying elastic weight consolidation allowed the quantum classifier to retain knowledge across all three tasks with an average accuracy reported as exceeding 92.3%. - For tasks involving engineered quantum data, the quantum classifier was reported to outperform a classical classifier with a comparable number of parameters. - The manuscript is an early, unedited version and will undergo further editing before final publication. Summary: The study presents an experimental approach to mitigating catastrophic forgetting in quantum learning systems and reports high average accuracy when using elastic weight consolidation. The authors also report a performance advantage for the quantum model over a similarly sized classical model on quantum-engineered data. The manuscript is an unedited pre-publication version and final publication details are pending.