When an image classifier decides, we can usually only estimate why. For quantum and all-optical classifiers that act linearly on amplitude-encoded pictures up to the measurement, with no trained classical network before or after them, the decomposition needs no approximation: each class's probability above its no-information share splits exactly into contributions of the encoded amplitudes, shown as Born evidence maps (BEM) over the picture and measurable on the hardware that classifies. For the quadratic form behind each class probability it equals Integrated Gradients from a zero baseline, in closed form. Deleting the 5 % of pixels with the highest evidence drops accuracy from 73 % to 9 %, as occlusion does with 1,024 extra runs per picture, whereas gradient saliency removes much less. One architecture serves both: a Fourier-plane mask followed by a brickwall circuit on qubits or by phase plates on a lens bench, read by a Born headroom readout with no trained parameter; centring each picture on its own mean lets dark features count as much as bright ones. In simulation, four phase plates at a single wavelength reach 72.3 % on Caltech-10 colour photographs (96.8 % on MNIST, 86.7 % on Fashion-MNIST, 47.9 % on colour CIFAR-10). On IBM's ibm_kingston, with the mask applied before approximate loading, 300 Caltech-10 test pictures were classified at 37.7 % against 10 % by chance (56.3 % in exact simulation), and 2x2 maps of ten correctly and confidently classified pictures, estimated on the chip, agreed with the exact ones (r = 0.77). The maps also revealed a shortcut in the data.
AI Summary
The paper asks whether quantum and optical image classifiers that are linear in encoded amplitudes until measurement admit exact, hardware-measurable input attributions. It defines Born evidence maps $e_g(x)=a_x\operatorname{Re}[A^\dagger(\Pi_g-b_gI)Aa]_x/Z$, satisfying $\sum_x e_g(x)=P_g-b_g$, where $Z=\|Aa\|^2$ and $b_g$ is a fixed class-share reference. These equal zero-baseline Integrated Gradients and Shapley attributions of the quadratic numerator with transmission held fixed—not attributions through the raw-pixel encoder or final nonlinear readout. A polarization identity estimates block evidence from whole-input, block-only and complement measurements.
The classifiers combine per-image mean-centred amplitude encoding, a learned Fourier mask, qubit circuits or optical phase plates, and a fixed-temperature headroom readout. Four simulated phase plates achieve $72.3\pm0.7\%$ on Caltech-10, $96.8\%$ on MNIST and $47.9\%$ on colour CIFAR-10. In a single-seed, 300-image deletion test, removing the highest-evidence 5% of pixels reduces optical accuracy from 73.0% to 8.7%, comparable to occlusion. On IBM hardware, 300 Caltech-10 images yield 37.7% accuracy versus 56.3% ideally and 41.3% with approximate loading alone. Hardware $2\times2$ maps for ten selected, correctly classified images correlate with ideal maps at $r=0.77$.
Limitations include classical mask preprocessing and approximate loading on the chip, simulation-only optical validation, unstable single-pixel maps across seeds, and classification accuracy below electronic CNNs. Caltech-10 preprocessing used test-set explanations, leaving no untouched holdout. No quantum or interference advantage is claimed.
Likely largely AI-generatedAI use disclosedIncrementalStrong assumptions
An AI assistant (Claude, Anthropic) was used throughout this work: to write and run simulation and analysis code, to prepare and check the hardware jobs, to draw the figures, and to draft and revise the text.
Error found: The abstract's unqualified reference to class probabilities obscures that Born evidence maps decompose the Born group mass P_g-b_g, not the classifier's reported softmax probability hat p_g.
For lossy optical classifiers, b_g=|S_g|/N is generally only a fixed anchor rather than the actual conditional class mass of a maximally mixed, no-information input.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
Comments
30 pages, 13 figures, 14 tables. Code, trained models and the IBM job records: https://github.com/stein-research/seeing-why
When an image classifier decides, we can usually only estimate why. For quantum and all-optical classifiers that act linearly on amplitude-encoded pictures up to the measurement, with no trained classical network before or after them, the decomposition needs no approximation: each class's probability above its no-information share splits exactly into contributions of the encoded amplitudes, shown as Born evidence maps (BEM) over the picture and measurable on the hardware that classifies. For the quadratic form behind each class probability it equals Integrated Gradients from a zero baseline, in closed form. Deleting the 5 % of pixels with the highest evidence drops accuracy from 73 % to 9 %, as occlusion does with 1,024 extra runs per picture, whereas gradient saliency removes much less. One architecture serves both: a Fourier-plane mask followed by a brickwall circuit on qubits or by phase plates on a lens bench, read by a Born headroom readout with no trained parameter; centring each picture on its own mean lets dark features count as much as bright ones. In simulation, four phase plates at a single wavelength reach 72.3 % on Caltech-10 colour photographs (96.8 % on MNIST, 86.7 % on Fashion-MNIST, 47.9 % on colour CIFAR-10). On IBM's ibm_kingston, with the mask applied before approximate loading, 300 Caltech-10 test pictures were classified at 37.7 % against 10 % by chance (56.3 % in exact simulation), and 2x2 maps of ten correctly and confidently classified pictures, estimated on the chip, agreed with the exact ones (r = 0.77). The maps also revealed a shortcut in the data.
The paper asks whether quantum and optical image classifiers that are linear in encoded amplitudes until measurement admit exact, hardware-measurable input attributions. It defines Born evidence maps eg(x)=axRe[A†(Πg−bgI)Aa]x/Z, satisfying ∑xeg(x)=Pg−bg, where Z=∥Aa∥2 and bg is a fixed class-share reference. These equal zero-baseline Integrated Gradients and Shapley attributions of the quadratic numerator with transmission held fixed—not attributions through the raw-pixel encoder or final nonlinear readout. A polarization identity estimates block evidence from whole-input, block-only and complement measurements.
The classifiers combine per-image mean-centred amplitude encoding, a learned Fourier mask, qubit circuits or optical phase plates, and a fixed-temperature headroom readout. Four simulated phase plates achieve 72.3±0.7% on Caltech-10, 96.8% on MNIST and 47.9% on colour CIFAR-10. In a single-seed, 300-image deletion test, removing the highest-evidence 5% of pixels reduces optical accuracy from 73.0% to 8.7%, comparable to occlusion. On IBM hardware, 300 Caltech-10 images yield 37.7% accuracy versus 56.3% ideally and 41.3% with approximate loading alone. Hardware 2×2 maps for ten selected, correctly classified images correlate with ideal maps at r=0.77.
Limitations include classical mask preprocessing and approximate loading on the chip, simulation-only optical validation, unstable single-pixel maps across seeds, and classification accuracy below electronic CNNs. Caltech-10 preprocessing used test-set explanations, leaving no untouched holdout. No quantum or interference advantage is claimed.
When a language model answers, we can usually only estimate why. We build a transformer in which every nonlinearity is a quadratic detection - an interference, a balanced reading or a camera - and the output is a Born probability: attention ||^2, the form of the Born rule, normalised per row, a measured product for the feed-forward nonlinearity, and a Born readout; between detections, classical control rescales and re-encodes. With its denominators held at the input, each class probability is a polynomial in the word amplitudes, so its integrated-gradient evidence is computed by a finite quadrature with no quadrature error: the words' contributions sum to the class probability above its no-information share to 10^-13. On e-SNLI natural language inference it reaches 77.45 +- 0.91 % accuracy, level with a same-size softmax transformer (76.62 +- 0.36 %) and an order-free bag of embeddings (77.79 %). Exactness buys the sum, not the best ranking. Deleting the word its evidence ranks first changes the answer in 44 % of pairs, of the 53 % some deletion changes, against 31-32 % for the softmax model's integrated gradients; the model's own gradient x input does as well (43 %), as does a bag of embeddings' (43 %, seed 0), and leave-one-out does better. Holding the denominators helps the low end: the five words ranked last change the answer in 1 %, against 6 % free. The same weights compile exactly to simulated optoelectronics, and their attention to qubits. On ibm_kingston, reading only the attention, the device gave the model's answers in all three Hadamard-test runs (the two shortest pairs and one without its deciding word) and both compute-uncompute runs (one circuit per overlap); uniform attention answers otherwise on two of the three inputs. With the device's row sums the evidence correlated with the exact one at 0.997 and 0.962 by Hadamard tests; with compute-uncompute, 0.971 and 0.48.
Introduction: A variational quantum classifier realises a function whose Fourier spectrum is fixed by the circuit and whose coefficients must be found by an optimiser. We study a readout borrowed from quantum phase estimation (QPE): a data-dependent operator acts on a target register through controlled powers, and the class label is read from an estimation register of m qubits after an inverse quantum Fourier transform (QFT). Materials and methods: We derive the exact readout distribution—a Fejér-weighted expansion in the moments of the operator—and characterise its Fourier spectrum for two single-qubit blocks: H ⋅ P h ⋅ R Y (family A) and a universal U 2 block (family B). All analytical statements are verified against a state-vector simulator, including attainment of the predicted spectrum. Three ablation controls alter one component each at an unchanged parameter count: the encoding QFT is removed, the exponential ladder of controlled powers is replaced by single powers, and the inverse QFT is replaced by independent Hadamard tests. The benchmark comprises 7680 training runs of 24 circuits on three tasks, with one generated dataset or one cross-validation fold as the unit of inference (twenty per sample size), exact signed-rank tests, and the corrected resampled t -test for the overlapping folds. Results: The support of every readout probability lies in parity-locked lattice cubes of maximal harmonic order 2 m + 1 . The maximal order is attained at every register size, in every feature and every outcome, except for family A with one target qubit, where the two highest orders cancel identically; the full support is attained at every register size used here ( t ≤ 4 , m ≤ 3 ), and the band is tapered towards its edge. Removing the encoding QFT costs family B 0.144 and 0.082 in test accuracy on the two-feature tasks, and replacing the power ladder by single powers costs 0.086 – 0.225 wherever m ≥ 2 ; replacing the inverse QFT by Hadamard tests has no detectable effect. The architecture leads the independently parameterised reference circuits only on the three-class task ( + 0.119 , p < 0.0001 ), where a parameter-tied re-uploading circuit with 12 rather than 20 parameters is not resolved from it ( + 0.023 , 95% confidence interval (CI)
We perform image classification on the Honda Scenes Dataset on Quantinuum's H-2 and IBM's Heron chips utilizing up to 72 qubits and thousands of two-qubit gates. For data loading, we extend the hierarchical learning to the task of approximate amplitude encoding and block amplitude encoding for commercially relevant images up to 2 million pixels. Hierarchical learning enables the training of variational circuits with shallow enough resources to fit within the classification pipeline. For comparison, we also study how classifier performance is affected by using piecewise angle encoding. At the end of the VQC, we employ a fully-connected layer between measured qubits and the output classes. Some deployed models are able to achieve above 90\% accuracy even on test images. In comparing with classical models, we find we are able to achieve close to state of the art accuracy with relatively few parameters. These results constitute the largest quantum experiment for image classification to date.
Image-based quantum machine learning usually follows a hybrid pipeline: a classical encoder compresses an image, and a quantum circuit then acts as a kernel, feature map, trainable classifier, or measurement-based readout. Yet it remains unclear which of these roles contributes useful predictive structure because prior studies often change the input representation, classical baseline, and measurement budget simultaneously. We present a controlled benchmark spanning five roles: quantum-kernel geometry, fixed product and entangled features, trainable data-reuploading representations, shot-aware probability readout, and task-adaptive Pauli measurement. Six quantum-derived methods, six matched classical controls, and two unmatched neural references are evaluated on MNIST, Fashion-MNIST, CIFAR-10, and BreastMNIST using a shared frozen ResNet-18 and train-only PCA interface. We assess exact and finite-shot inference, predictive performance, calibration, selective risk, resource requirements, and statistical reliability. PdrQC-Matched was the strongest quantum-derived method on all four datasets. C-RBF was better on Fashion-MNIST and CIFAR-10 by 2.19 and 3.50 accuracy points, while BreastMNIST showed no detectable difference. Despite lower exact accuracy, PdrQC-Matched achieved significantly lower Brier score and AURC than C-RBF on Fashion-MNIST and CIFAR-10. Finite-shot results showed no automatic benefit from entanglement or shot-aware training. Overall, image-QML conclusions depend on quantum role, measurement budget, classical control, and evaluation metric.
Quantum machine learning on real noisy intermediate-scale quantum (NISQ) hardware has remained largely confined to binary or few-class tasks, limited by the cost of on-hardware training and the underuse of large devices at inference. We present a unified framework that classifies ten-class MNIST end-to-end on a 127-qubit IBM Eagle processor, with three central contributions. First, a two-phase protocol decouples a gradient-based classical optimization of the encoder and readout from a gradient-free optimization of the quantum parameters, removing the parameter-shift gradient cost that makes on-hardware training impractical. Second, we introduce Quantum Multi-Programming to a trained quantum classifier for the first time, packing multiple circuit copies onto one device to deliver parallel inference at no mean-accuracy cost while cutting quantum-processor job submissions proportionally. Third, a controlled comparison shows that on-hardware fine-tuning yields no measurable accuracy gain, motivating a practical NISQ workflow: train on a classical simulator and reserve the hardware for inference only. Benchmarked against a matched-capacity classical network, the quantum module shows no per-parameter accuracy advantage at this scale; we therefore frame the work as a feasibility-and-workflow demonstration for multi-class quantum image classification on current hardware.