Determine whether SamplerQNN postprocessing caused the reported transfer-learning performance gap

Determine whether the SamplerQNN postprocessing data-loss mechanism caused the reported hardware-performance gap in the hybrid classical-quantum transfer-learning experiments of Martin-Perez et al., which used four virtual qubits on the 133-qubit ibm_torino backend through SamplerV2.

Background

The paper identifies a potential data-loss mechanism in SamplerQNN: an integer-range filter can discard valid measurement shots when hardware returns bit-strings spanning many physical qubits rather than only the virtual qubits. The authors note that Martin-Perez et al. used a four-qubit SamplerQNN with SamplerV2 on ibm_torino, a 133-qubit backend, matching the conditions under which substantial shot loss can occur.

Martin-Perez et al. reported different hardware-versus-simulation performance changes for several classical backbones, including a 19-percentage-point drop for EfficientNet-B0. Although that gap was attributed to transpilation overhead and stochastic gradients, the present paper states that the available information is insufficient to determine whether the SamplerQNN postprocessing defect contributed to it; resolving the issue would require access to the raw measurement data.

References

We note that we cannot confirm whether the behaviour we document caused the gap in Ref. [18] without inspecting their raw measurement data, but the setup (4 virtual qubits on a 133-qubit backend via SamplerQNN and SamplerV2) matches the conditions under which the behaviour arises.

— Impact of Data Loss in Postprocessing on Training and Inference of Quantum Neural Networks  (2609.05060 - Panambalom et al., 4 Sep 2026) in Section V, discussion of Ref. [18]