跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.01993 · v3

Seeing Why: Exact Evidence Maps for Quantum and Optical Image Classifiers

Teerawat TongloyRangsan Jomtarak
quant-phcs.LGphysics.optics
首次提交 2026-11-01 23:27
版本更新 2026-11-06 00:21
雷达观测 2026-11-06 12:09
arXiv ↗PDF ↗

摘要

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, $P_g - b_g$, 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 $g\times g$ map takes $1 + 2g^2$ runs. Deleting the 5% of pixels with the highest evidence drops accuracy from 73% to 9% (one seed), as occlusion and gradient $\times$ 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 $2\times2$ maps of the filtered input of 30 test pictures drawn at random in advance, measured on the chip, agreed with the exact ones ($r = 0.88$; 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.

作者备注：38 pages, 13 figures, 16 tables. Code, models and IBM job records: https://github.com/stein-research/seeing-why (v1.2, doi:10.5281/zenodo.23984169). v3: chip maps of 30 pictures drawn at random in advance (r = 0.88); a second model, with no mask, on the chip; reviewer checks (nulls, model randomisation, ROAR, calibration), most in a new appendix; corrections (loading circuit 44 CNOTs)

查看 arXiv 原文 ↗阅读 PDF ↗

AI 论文解读

基于摘要 · 非全文审阅

作者提出一种可在量子芯片或光学实验台上直接测量的精确证据图（BEM），把线性量子/全光图像分类器的类别概率超出无信息份额的部分，精确分解到编码振幅上。

正在读取账号与研究记录…

版本差异

仅对比实际观测的元数据，不代表 PDF 正文或实验变化。版本可能不连续；带删除线为移除，下划线为新增。

摘要变化

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, $P_g - b_g$, splits exactly into contributions of the encoded amplitudes, shown as Born evidence maps (BEM) over the picture. 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 $g\times g$ map takes $1 + 2g^2$ runs. Deleting the 5% of pixels with the highest evidence drops accuracy from 73% to 9% (one seed), as occlusion does with 1,024and extragradient runs$\times$ perinput picturedo, whereaswhile gradientplain saliencygradients removesremove 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 a singleone wavelength reach 72.3% on Caltech-10 colour photographs. On IBM's ibm_kingston, with the mask appliedcomputed classically before approximate loading, 300 test pictures were classified at 37.7% against 10% by chance (56.3% in exact simulation), and $2\times2$ maps of tenthe correctlyfiltered classifiedinput pictures,of one30 pertest classpictures chosendrawn byat arandom statedin ruleadvance, measured on the chip, agreed with the exact ones ($r = 0.7788$; 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.

作者备注变化

3038 pages, 13 figures, 1416 tables. Code, trained models and the IBM job records: https://github.com/stein-research/seeing-why (archived:v1.2, doi:10.5281/zenodo.2390635723984169). v2v3: abstractchip andmaps introductionof clarified30 pictures drawn at random in advance (knownr identity= vs0.88); newa measurement;second Caltech-10model, resultswith markedno exploratory)mask, correspondingon authorthe chip; reviewer checks (nulls, authors'model ORCIDsrandomisation, ZenodoROAR, DOIcalibration), clearermost Figs.in 9a andnew 12appendix; resultscorrections unchanged(loading circuit 44 CNOTs)

观测到的版本

相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。

v3 · 2026-11-06 00:21 · 最新观测版本

Seeing Why: Exact Evidence Maps for Quantum and Optical Image Classifiers

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, $P_g - b_g$, 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 $g\times g$ map takes $1 + 2g^2$ runs. Deleting the 5% of pixels with the highest evidence drops accuracy from 73% to 9% (one seed), as occlusion and gradient $\times$ 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 $2\times2$ maps of the filtered input of 30 test pictures drawn at random in advance, measured on the chip, agreed with the exact ones ($r = 0.88$; 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.

查看此版本解读 →查看此版本原文 ↗
v2 · 2026-11-03 20:24

Seeing Why: Exact Evidence Maps for Quantum and Optical Image Classifiers

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, $P_g - b_g$, splits exactly into contributions of the encoded amplitudes, shown as Born evidence maps (BEM) over the picture. 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 from camera frames on a lens bench, with no gradient circuit and no light sent backwards; a $g\times g$ map takes $1 + 2g^2$ runs. 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. 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 a single wavelength reach 72.3% on Caltech-10 colour photographs. On IBM's ibm_kingston, with the mask applied before approximate loading, 300 test pictures were classified at 37.7% against 10% by chance (56.3% in exact simulation), and $2\times2$ maps of ten correctly classified pictures, one per class chosen by a stated rule, measured on the chip, agreed with the exact ones ($r = 0.77$). 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.

查看此版本解读 →查看此版本原文 ↗