Impact of Data Loss in Postprocessing on Training and Inference of Quantum Neural Networks

Soraya V. Panambalom; Edoardo Altamura; Nick Chancellor; Jonte R. Hance

quant-phSep 4, 2026

Impact of Data Loss in Postprocessing on Training and Inference of Quantum Neural Networks

Authors: Soraya V. Panambalom, Edoardo Altamura, Nick Chancellor, Jonte R. Hance

Organizations: Quantum Group, School of Computing, Newcastle University, 1 Science Square, Newcastle upon Tyne, NE4 5TG, UK · National Quantum Computing Centre, Didcot, OX11 0QX, UK · Yusuf Hamied Department of Chemistry, University of Cambridge, Lensfield Road, Cambridge CB2 1EW, UK

Abstract

As quantum hardware scales to larger devices, the classical software layers that interface with it must evolve in step. Postprocessing routines developed and tested primarily in simulator settings can encode assumptions that no longer hold on utility-scale devices, leading to data loss that can be difficult to detect from high-level model outputs alone. We present a case study of SamplerQNN, the sampling-based quantum neural network class in the Qiskit Machine Learning library. Here, the postprocessing method applies a filter that assumes measurement bit-strings are in virtual qubit space. On our quantum hardware runs, where bit-strings span over 100 physical qubits, this filter led to the loss of 85 to 99.6% of valid measurement shots, depending on the transpiler's qubit placement. The resulting probability vector is unnormalised, allowing distorted prediction and loss values to propagate through the model without an API-level warning. We demonstrate the impact across five experiments on two IBM backends: for inference, accuracy drops from 0.94 to 0.39 on the same raw measurements; for training, the loss signal is compressed by 22 to 27 $\times$ , substantially reducing the sensitivity of the optimiser to the objective landscape. The behaviour arises in all released versions of the library (0.8.4 to 0.9.0). We implemented a layout-based marginalisation fix, merged into the GitHub codebase as Pull Request #1041, that makes SamplerQNN postprocessing forward-compatible with current and upcoming hardware.

Figures & tables

Model	Q	Backend	Positions	Surv.
CS	4	ibm_kingston	[0,1,2,3]	15.2%
CS	7	ibm_torino	[0,1,2,3,4,5,6]	4.1%
CS (SI+FE)	4	ibm_kingston	[108,109,110,118]	0.4%
VQC (ZZ)	4	ibm_torino	[61,62,54,60]	0.2%
VQC (Z)	7	ibm_torino	[0,1,2,3,4,5,6]	5.2%

TABLE I: Physical qubit positions assigned by the transpiler and percentage of measurement shots surviving the SamplerQNN filter for each experiment.

Model	Qubits	Method A	Method B	Method C
CS	4	0.57	0.94	0.94
CS	7	0.63	0.92	0.89
CS (SI+FE)	4	0.56	0.94	0.39
VQC (ZZ)	4	0.44	0.54	0.39
VQC (Z)	7	0.45	0.59	0.61

TABLE II: Accuracy comparison of three postprocessing methods on the same raw hardware measurements. CS: Classifier Sampler. SI+FE: Small-Angle Initialisation with Full Entanglement. Methods A, B, and C are defined in Section III-A .

Model	Qubits	Method A	Method B	Method C
CS	4	0.56	0.93	0.93
CS	7	0.60	0.91	0.88
CS (SI+FE)	4	0.56	0.93	0.37
VQC (ZZ)	4	0.43	0.50	0.31
VQC (Z)	7	0.45	0.53	0.57

TABLE III: F1 score (macro) comparison of three postprocessing methods on the same raw hardware measurements. CS: Classifier Sampler. SI+FE: Small-Angle Initialisation with Full Entanglement. Methods A, B, and C are defined in Section III-A .

	Met. C (filter)			Met. B (marginal.)
Class	Prec.	Rec.	F1	Prec.	Rec.	F1
Malignant	0.34	0.74	0.47	0.95	0.88	0.91
Benign	0.54	0.18	0.27	0.93	0.97	0.95
Accuracy	0.39			0.94

TABLE IV: Per-class classification metrics computed from the same raw hardware measurements. Method C retains only 16 of 4096 shots (0.4%); Method B uses all 4096.

Model	Iter.	A	B	C	Surv.
4Q	1 (clean)	0.50	0.33	0.012	3.7%
	2 (affected)	0.50	0.42	0.016	3.7%
	3 (affected)	0.50	0.33	0.012	3.7%
	4 (affected)	0.50	0.41	0.016	4.0%
	5 (affected)	0.50	0.34	0.014	4.0%
7Q	1 (clean)	0.50	0.41	0.019	4.6%

TABLE V: Loss recomputed from raw measurements at each hardware iteration. Method C is what the optimiser received; Method B is the correct loss. Iteration 1 uses clean simulator-trained weights; iterations 2 to 5 use weights affected by prior uncorrected updates.

Fig. 1: Loss for 5 on-device fine-tuning iterations, recomputed from raw measurements. Method B (correct marginalisation) reveals the actual loss landscape; Method C (original filter) shows what COBYLA received. The star marks iteration 1 (clean simulator-trained weights); the shaded region marks iterations where the weights had already been affected by prior uncorrected updates.

Model	Iter.	$\Delta w_{0}$	$\Delta w_{1}$	$\Delta w_{2}$	$\Delta w_{3}$	# changed	Loss C
4Q (12 weights)	1 (clean)	0	0	0	0	0	0.012
	2	+1.0	0	0	0	1	0.016
	3	0	+1.0	0	0	1	0.012
	4	0	+1.0	+1.0	0	2	0.016
	5	0	+1.0	0	+1.0	2	0.014
	inference	0	+1.0	0	0	1	$-$

TABLE VI: Weight drift relative to clean simulator-trained values at each iteration. All non-zero changes are exactly

\pm

1.0 rad (= rhobeg ), confirming simplex probing. Loss C is the value the optimiser received and used to decide whether to keep or revert each perturbation.

Quantum Neural Networks

Quantum neural networks use quantum circuits and measurements to learn representations, make predictions, and model complex data.

See all 135 papers

Scalable On-Hardware Training of Quantum Neural Networks and Application to Clinical Data Imputation
Jun 2, 2026Natansh Mathur, Panagiotis Kl. Barkoutsos, Masako Yamada +2Hybrid Quantum-Classical ML Quantum Neural Networks
IRIF, CNRS and Université Paris Cité and QC Ware, France · QC Ware, France · IonQ +2
Modeling quantum neural network gradient with reinforcement learning
Sep 25, 2026Nhan Trong Luu, Duong Trung Luu, Nam Ngoc Pham +1Quantum Neural Networks Quantum Circuit Trainability
College of Information and Communication Technology, Can Tho University, Can Tho, Vietnam · College of Engineering and Computer Science, VinUniversity, Hanoi, Vietnam · Center for Digital Transformation and Communications, Can Tho University, Can Tho, Vietnam +1
Eliminating Vendor Lock-In in Quantum Machine Learning via Framework-Agnostic Neural Networks
Apr 6, 2026Poornima Kumaresan, Shwetha Singaravelu, Lakshmi Rajendran +1Quantum Neural Networks Hybrid Quantum-Classical ML
Intrinsic Lab, Centre for Sensors, Instrumentation and Cyber-Physical System Engineering (SeNSE), Indian Institute of Technology Delhi, New Delhi 110016, India

Jun 2, 2026J/K move · Enter open · S save

Scalable On-Hardware Training of Quantum Neural Networks and Application to Clinical Data Imputation

Natansh Mathur, Panagiotis Kl. Barkoutsos, Masako Yamada, Martin Roetteler+1

IRIF, CNRS and Université Paris Cité and QC Ware, France · QC Ware, France · IonQ+2

Hybrid Quantum-Classical ML Quantum Neural Networks

Open paper AskPDF arXiv

Training quantum neural networks (QNNs) on quantum hardware is currently bottlenecked by the cost of gradient estimation: standard parameter-shift methods require a number of circuit evaluations that grows quadratically with the number of trainable parameters, making hardware-based optimisation impractical beyond small system sizes. In this work, we introduce a training framework that reduces this cost to logarithmic in the number of qubits, making gradient-based QNN optimisation feasible on near-term hardware at increasing scales. Our framework combines three co-designed ingredients: (i) a structured, subspace-preserving Butterfly circuit architecture with

O(n \log n)

parameters and logarithmic depth; (ii) a layer-wise training strategy that confines on-hardware optimisation to one small, well-structured layer at a time; and (iii) a parallelised parameter-shift rule that exploits the commuting structure within each Butterfly layer to extract all gradients in a constant number of circuit executions. Together these reduce the number of distinct circuit evaluations per optimisation step from

O(n^2)

O(\log n)

. We validate the framework on clinical data imputation using the MIMIC-III electronic health record dataset, a demanding benchmark sensitive to optimisation instability and model variance. Hybrid classical-quantum models are trained directly on IonQ Forte Enterprise trapped-ion hardware at 16 qubits without performance degradation relative to ideal or noisy simulation and via tensor-network simulation at 32 qubits, with 32-qubit inference executed on hardware. The resulting models match or exceed strong classical neural baselines in downstream patient survival prediction while exhibiting reduced variance across runs, demonstrating that the proposed framework enables practical, scalable QNN training under realistic hardware constraints.

Abstract

Quantum Neural Networks

Explore similar work

Scalable On-Hardware Training of Quantum Neural Networks and Application to Clinical Data Imputation

Modeling quantum neural network gradient with reinforcement learning

Eliminating Vendor Lock-In in Quantum Machine Learning via Framework-Agnostic Neural Networks