Seeing Why: Exact Evidence Maps for Quantum and Optical Image Classifiers
Organizations: School of Integrated Innovative Technology, King Mongkut’s Institute of Technology Ladkrabang, Bangkok, Thailand · Faculty of Science and Technology, Suan Dusit University, Bangkok, Thailand
Abstract
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, each class's Born probability above its no-information share, , splits exactly into contributions of the encoded amplitudes, shown as Born evidence maps (BEM). The split itself follows from known results on quadratic forms: with the transmitted light held fixed, it is what Integrated Gradients from a zero baseline and the Shapley value give. What we add is a way to measure it on the hardware that classifies, from class probabilities on a quantum chip or, in simulation, from camera frames on a lens bench, with no gradient circuit and no light sent backwards; a map takes runs. Deleting the 5% of pixels with the highest evidence drops accuracy from 73% to 9% (one seed), as occlusion and gradient input do, while plain gradients remove much less. The classifiers share one architecture, a Fourier-plane mask followed by a brickwall circuit on qubits or by phase plates on a lens bench, read by a fixed 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 one wavelength reach 72.3% on Caltech-10 colour photographs. On IBM's ibm_kingston, with the mask computed classically before approximate loading, 300 test pictures were classified at 37.7% against 10% chance (56.3% in exact simulation), and maps of the filtered input of 30 test pictures drawn at random in advance, measured on the chip, agreed with the exact ones (; standard estimator 0.82). The maps also revealed a shortcut in the data; as the trim that removed it was chosen on test pictures, the Caltech-10 results are exploratory.
Figures & tables
| row | model | params | accuracy (%) | ROC-AUC | transm. |
| A | phase plates, plain amplitudes | 4,096 | 0.419 | ||
| B | + mean-centred encoding | 4,096 | 0.405 | ||
| C | + lens mask | 4,614 | 0.044 | ||
| D | + phase in the mask | 5,124 | 0.042 | ||
| E | D, pixel-aware plates | 5,124 | 0.020 | ||
| F | E + light-keeping loss | 5,124 | 0.038 |
| model (Caltech-10) | parameters | accuracy (%) |
|---|---|---|
| qubit brickwall + mask, phase, mean-centred encoding (D ′ ) | 1,460 | |
| two phase plates of | 1,540 | 46.7 |
| four phase plates of | 2,052 | 63.3 |
| two phase plates of | 3,076 | 59.7 |
| four phase plates of (row D) | 5,124 | |
| two phase plates of | 9,220 | 70.7 |
| plain amplitudes | mean-centred | |||||
|---|---|---|---|---|---|---|
| dataset | pictures | train / test | accuracy (%) | ROC-AUC | accuracy (%) | ROC-AUC |
| MNIST | grey | 58k / 10k | ||||
| Fashion-MNIST | grey | 58k / 10k | ||||
| CIFAR-10 | grey | 48k / 10k | ||||
| CIFAR-10 | colour | 48k / 10k | ||||
| Caltech-10 | colour | 2,644 / 300 | ||||
| plates (row D), 73.0 % | qubits (D ′ ), 56.3 % | |||||||
|---|---|---|---|---|---|---|---|---|
| ranked by deleted: | 2 % | 5 % | 10 % | 20 % | 2 % | 5 % | 10 % | 20 % |
| evidence (ours) | 36.3 | 8.7 | 1.0 | 0.3 | 20.7 | 4.7 | 0.0 | 0.0 |
| occlusion | 36.0 | 9.3 | 1.7 | 0.3 | 21.0 | 4.3 | 0.3 | 0.0 |
| gradient input | 37.0 | 9.0 | 1.3 | 0.3 | 20.3 | 4.3 | 0.3 | 0.0 |
| gradient input, | 31.3 | 5.0 | 0.7 | 0.3 | 16.0 | 2.7 | 0.0 | 0.0 |
| gradient | 68.7 | 65.0 | 60.0 | 53.0 | 55.0 | 52.3 | 48.3 | 45.0 |
| answers kept (%) after deleting | kept after 2 % | lead | |||||
|---|---|---|---|---|---|---|---|
| 2 % | 5 % | 10 % | 20 % | right | wrong | before after 5 % | |
| plates (row D) | 39.3 | 9.3 | 1.0 | 0.3 | 49.8 | 11.1 | |
| qubits (D ′ ) | 26.0 | 4.7 | 0.0 | 0.0 | 36.7 | 12.2 | |
| row (Table 1 ) | input intensity / area | positive evidence / area | its peak inside |
|---|---|---|---|
| A plates, plain | 1.02 | 1.04 | 0.49 |
| B + mean-centred | 1.42 | 1.32 | 0.68 |
| D full optical model | 1.42 | 1.37 | 0.71 |
| A ′ qubit, plain | 1.02 | 1.00 | 0.42 |
| D ′ qubit, full front end | 1.42 | 1.26 | 0.62 |
| maps against the exact ones | correlation | centred | same sign | strongest block |
|---|---|---|---|---|
| 30 pictures drawn at random (120 blocks) | ||||
| ibm_kingston, Eq. ( 6 ) | 0.88 (0.84–0.91) | 0.93 | 88 % | 23 / 30 |
| ibm_kingston, Eq. ( 5 ) | 0.82 (0.76–0.87) | 0.92 | 84 % | 21 / 30 |
| noisy simulation, 4,000 shots, Eq. ( 6 ) | 0.90 | 0.95 | 82 % | |
| noisy simulation, 4,000 shots, Eq. ( 5 ) | 0.80 | 0.94 | 79 % | |
| 10 pictures chosen by a rule (40 blocks), earlier maps | ||||
| LightPipes bench (10 test pictures) | overlap (min) | same prediction |
|---|---|---|
| row E, 4f output passed on as it is | 0.997 (0.996) | 10 / 10 |
| row E, 4f output re-made into square pixels | 0.998 (0.996) | 10 / 10 |
| row E, glass plates ( ) crossed in 32 layers | 0.997 (0.9945) | 10 / 10 |
| row E, convolution propagator instead of the spectral one | 0.995 (0.993) | 8 / 10 |
| row D, point-sampled plates, 4f output as it is | 0.980 (0.971) | 7 / 10 |
| per-pixel map ( ) | airplane | leopard | ketch | ketch’s sum |
|---|---|---|---|---|
| correlation with our model’s exact map | ||||
| bench, no noise | 0.94 | 0.90 | 0.91 | 0.038 |
| photons per frame | 0.94 | 0.88 | 0.90 | |
| photons per frame | 0.60 | 0.43 | 0.46 | |
| correlation with the flawless bench’s map | ||||
| laser drift 0.1 % per frame | 0.996 | 0.998 | 0.984 | 0.042 |
| flaw | tolerated | breaks |
|---|---|---|
| picture modulator levels | 3–4 bits | 2 bits: 68.7 % |
| mask modulator levels | 5 bits | 4 bits: 70.0 %; 3 bits: 72.0 %; 2 bits: 58.0 % |
| plate phase levels | 5–6 bits | 4 bits: 71.0 %; 2 bits: 44.0 % |
| plate sideways shift | 0.05 plate px (0.8 µm; 72.9 %) | 0.1 px: 69.2 %; 0.25 px: 51.2 %; 0.5 px: 24.7 % |
| plate rotation | 0.25 ∘ (71.1 %) | 0.5 ∘ : 66.0 %; 1 ∘ : 58.9 %; 2 ∘ : 41.9 % |
| mask off-centre | 0.25 mask px (2 µm; 71.7 %) | 0.5 px: 63.6 %; 1 px: 38.2 % |
| trained with shifts up to (plate, picture) | aligned | plate off 0.25 px | plate off 0.5 px |
|---|---|---|---|
| none | 68.7 | 54.0 | 23.1 |
| 0.25 plate px, 0.5 px | 66.7 | 58.3 | 32.7 |
| 0.5 plate px, 1 px | 65.3 | 60.9 | 46.6 |
| work | classical network? | data | reported result | explanation |
| qubits | ||||
| QCNN [ 4 ] | none | quantum states | phase recognition | |
| circuit-centric [ 5 ] | none | small benchmarks | simulation | |
| quantum kernels [ 7 ] | none | few features | superconducting chip | |
| transfer learning [ 6 ] | pretrained CNN in front | images | IBM and Rigetti chips | |
| QGrad-CAM [ 8 ] | CNN in front | MNIST (10), dogs vs cats, TIMIT speech | simulation | gradient (Grad-CAM) |
| all 300 pictures | both right | ||||
|---|---|---|---|---|---|
| model | pixels | blocks of 4 px | blocks of 8 px | blocks of 16 px | pixels |
| plates (row D) | 0.25 (0.16–0.35) | 0.69 (0.55–0.79) | 0.78 (0.63–0.88) | 0.89 (0.61–0.96) | 0.29 (0.19–0.40) |
| qubits (D ′ ) | 0.10 (0.02–0.18) | 0.26 (0.09–0.43) | 0.40 (0.16–0.63) | 0.56 ( 0.09–0.87) | 0.16 (0.09–0.27) |
| plates vs qubits | 0.09 (0.02–0.19) | 0.28 (0.07–0.48) | 0.37 (0.08–0.64) | 0.42 ( 0.16–0.82) | 0.15 (0.06–0.26) |
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| run | jobs | circuits | quantum s |
|---|---|---|---|
| test set, model D ′ (2026-10-31) | 10 | 300 | 340 |
| dbit0pohoj4c73b6str0 dbit2c6ne4hc73bf93b0 dbit2rmne4hc73bf9490 dbit3cohoj4c73b6t2og dbit3s6ne4hc73bf96m0 dbit4c8hoj4c73b6t4m0 dbit4tune4hc73bf98o0 dbit5d8hoj4c73b6t6pg dbit5t6ne4hc73bf9aug dbit6e5ep8hs73aohpsg | |||
| earlier maps, first run (2026-10-31) | 1 | 80 | 87 |
| dbitsbghoj4c73b6u5s0 | |||
| earlier maps, second run (2026-10-31) | 1 | 90 | 97 |
| dbitvl01h92s738469gg | |||
| hardware | |
| GPU, CPU, memory | NVIDIA GeForce RTX 5050 Laptop GPU (8 GB); Intel Core i5-13450HX; 16 GB |
| quantum computer | IBM ibm_kingston (Heron r2) through Qiskit Runtime |
| software | |
| Python 3.12.3 | |
| PyTorch 2.11 (CUDA 12.8) [ 32 ] | simulator, optics, training |
| NumPy 2.5, SciPy 1.18 | arrays, data |
| check | question | result |
|---|---|---|
| readout slope | does BHR’s accuracy rest on its slope alone? | 69.8–72.3 % for slopes 68–272 (row D) |
| calibration | are the readouts’ confidences calibrated? | BHR underconfident (ECE 0.10), 0.06 after temperature scaling |
| retraining on deleted pictures | can ROAR rank the maps? | no: the deleted pixels’ pattern alone gives the class (90–98 %) |
| reach of the mask’s kernel | how far does a block of the filtered state stand for the picture? | 66 % of the kernel’s energy on the pixel itself |
| steps of the map error | where does the chip’s map error come from? | preparation slope 1.09, noise 0.90, chip 0.82 |
| maps | finer maps on the chip | , at the shot-noise limit |
Explainable Image Classification
Explainable image classification develops ways to show which pixels, regions, concepts, or model features support an image classifier’s predictions.
See all 104 papers