Speaker
Description
Quantum machine learning is a promising paradigm for learning from limited data, a bottleneck in data scarce regimes like medical imaging. Quantum convolutional neural networks are especially attractive here, combining a hierarchical architecture, strong inductive bias, and a parameter count that grows logarithmically with system size. Their appeal rests on generalization bounds showing that a quantum model's error scales with its number of trainable gates rather than the Hilbert-space dimension, placing QCNNs in a sample-efficient regime. Recent work cautions, however, that such bounds cannot fully explain generalization: quantum models can fit randomized labels, and QCNNs are classically simulable on simple benchmarks.
We develop a hardware-compatible QCNN with mid-circuit measurement and classical feed-forward. On a binary handwritten-digit task, strong test performance emerges from as few as 10 training samples and the generalization error steadily falls as the data increases. At a matched 45-parameter budget the QCNN learns where an equally small classical network stays at chance, though an unconstrained 25,000-parameter baseline wins when data are plentiful. Transpiling amplitude and angle encoded circuits from 2×2 to 512×512 pixels exposes the scaling bottleneck: amplitude encoding stays qubit-efficient (18 qubits) but grows extremely deep, while angle encoding stays shallow but qubit-prohibitive (over 260,000 qubits). On the BreastMNIST benchmark, the QCNN does not surpass the unconstrained network yet learns above chance with far fewer parameters. Learning from few samples is thus attainable in practice, but scaling is limited by data encoding and hardware execution, not optimization; we read this as data efficiency, not quantum advantage.
| Topic | Quantum algorithms for machine learning applications |
|---|