How Much Reconstruction Does Quantum Machine Learning Need? Late Fusion of Independently Trained Quantum Subcircuits

As quantum neural networks (QNNs) scale beyond the capacity of individual devices, researchers use circuit cutting to partition them into smaller subcircuits. Traditionally, these subcircuits are recombined using quasiprobability reconstruction, which incurs an exponential sampling overhead. This paper investigates whether this expensive reconstruction step is strictly necessary for machine learning tasks or if a simpler, decision-level combination of independently trained subcircuits can achieve comparable performance.

The authors propose "late fusion," a technique adapted from multimodal deep learning. Instead of reconstructing the original quantum circuit, each subcircuit is trained and measured independently. A small, trainable classical head then combines these local measurement features into a final prediction. To explore the trade-off between this efficient fusion and exact reconstruction, the authors introduce a "quantumness dial" (Q), which allows for a tunable reconstruction budget, and a cut-entanglement diagnostic that predicts how much reconstruction a specific task requires based on the entanglement entropy across the cut.

• Efficiency: Late fusion avoids the exponential sampling overhead of reconstruction, maintaining a linear cost relative to the number of subcircuits.
• Accuracy: Across both synthetic and standard datasets, late fusion consistently matches the accuracy of full reconstruction within a 0.04 margin.
• Robustness: Because reconstruction relies on a signed sum of quasiprobabilities, it is highly sensitive to shot noise and device errors. Late fusion is significantly more robust, as it avoids this noise amplification.
• Diagnostic Utility: The authors demonstrate that the cut-entanglement diagnostic effectively predicts the operating point where fusion is sufficient versus where reconstruction is necessary, providing a clear boundary for when to apply each method.