Frank and Frank (2020) trained a convolutional neural network (CNN) on entropy-filtered image tiles of Rembrandt portraits and van Gogh landscapes and rendered its per-region verdicts as color-coded probability maps, correctly classifying several historically contested works and one modern forgery. They emphasized, however, that the method's evidence is opaque and causally ambiguous: the network 'resists reverse engineering', so the visual cues distinguishing an artist's hand remain uncharacterized, and a region classified as 'not by the master' could equally reflect a workshop assistant's contribution, a later restoration, retouching of the digitized photograph, or an unusual stylistic experiment by the master himself. Cross-genre transfer of the classifier was also left unverified, and the authors note that the extent of any such network's generalization is in principle unknowable, since an artist's full range of style and its limitless imitations always surpass the training distribution. In the subsequent literature (2021–2026), surface-topography machine learning, Vision-Transformer authentication, Raphael and Rubens attribution studies, and explainability toolkits for AI-generated art have improved discrimination accuracy, but none characterizes the cues a painting-attribution network relies on, none causally separates the competing causes of a negative attribution on real paintings, and cross-genre transfer remains untested. The open core — interpretable, cause-specific computational attribution — survives intact.

Background

In art history, 'attribution' is the judgment of who made a work, traditionally exercised by connoisseurship: the expert's eye built from long study of an artist's accepted oeuvre. For Rembrandt this is exceptionally hard because he ran an active workshop with roughly a hundred students and assistants over his career; pupils such as Govaert Flinck and Carel Fabritius copied and completed his compositions, so many canvases carry mixed hands, and attributions of famous works (e.g., 'Man with the Golden Helmet') have been debated for decades. Scientific techniques such as X-ray radiography, dendrochronology (dating a wooden panel by its tree rings), and pigment analysis answer when and from what a work was made, but not by whom in the stylistic sense.

Against this background, Frank and Frank (2020) proposed a computational approach. A convolutional neural network — a deep image classifier that learns its own visual features from labeled examples — was trained to distinguish tiles (square image crops) of Rembrandt portraits from tiles of portraits by his students, his 'school', and Dutch contemporaries. Only tiles whose 'image entropy' (a measure of visual diversity from information theory) met a threshold were examined, and each examined region received a probability of Rembrandt authorship, visualized as a four-color 'probability map' over the painting. The same pipeline, trained on van Gogh landscapes, correctly rejected a painting by the forger John Myatt. Notably, the diagnostic scale differed by artist: for Rembrandt, tiles small enough to confine analysis to brushstroke-level detail classified barely better than chance, whereas van Gogh's distinctiveness was most pronounced exactly at brushstroke scale — consistent with earlier work by Li, Yao, Hendriks and Wang (2012), whose automated brushstroke extraction separated van Gogh from contemporaries using hand-designed features.

The authors confronted the method's evidentiary limits. For ordinary object recognition, tools such as Grad-CAM (gradient-weighted class activation mapping, which highlights the image regions most responsible for a CNN's decision) can visualize intermediate features — a cat's eyes and ears. But what separates an artist from imitators 'is infinitely more subtle, and the predictively significant features do not cohere into any sort of unifying theme', so the network 'resists reverse engineering'. Worse, a blue ('not-Rembrandt') region is causally underdetermined: it may record a workshop assistant's hand, heavy later restoration, retouching of the digital photograph (e.g., to remove glare), or simply an unusual manner by the master that falls outside the training set; these possibilities 'coexist, with varying degrees of likelihood, until resolved by records or authoritative scholarship — if ever'. A single cross-genre test (van Gogh's still life 'Three Pairs of Shoes', classified from a landscape-trained network, at a much lower probability than his landscapes) was flagged by the authors as carrying 'intrinsic risks' left for future work. They also stressed an epistemic bound: because an artist's full stylistic range and the limitless ways it can be imitated always exceed any training corpus, the generalization limits of such a classifier are unknowable, and any authorship hypothesis it supports must be qualified accordingly.

Later work has expanded the discrimination toolkit rather than the evidentiary one. Ji et al. (2021) trained CNN ensembles on the three-dimensional surface topography of paintings (measured by confocal optical profilometry, an instrument that maps surface height at micron scale) and found that very short length scales — down to twice a paintbrush bristle's diameter — distinguish individual artists' hands, explicitly framing workshop-hand and conservation-intervention separation as future work. Schaerf, Popovici and Postma (2023) compared Vision Transformers (attention-based image classifiers) with CNNs for van Gogh authentication, finding imitation detection to be the persistently hard part. Ugail et al. (2023) combined image features from ResNet50 (a standard 50-layer convolutional network), a support vector machine classifier, and edge-detection analysis for Raphael attribution. Afifi et al. (2025) applied tile-level CNN ensembles to Rubens and his workshop, publishing confidence and uncertainty maps that still must be interpreted narratively. Generic saliency and occlusion methods (highlighting regions whose masking changes the classifier's output) predate and postdate the source paper, including the ArtBrain toolkit's multi-class saliency maps for AI-generated art attribution, but they localize where a network looks, not what painterly cues identify a historical master's hand. The question the source paper left — can computational attribution produce evidence whose meaning and cause can be established, rather than an opaque probability — therefore remains the live one.

Problem Statement

The problem, as posed by the source's own limitations, is to turn computational artwork attribution from an opaque probability into interpretable, cause-specific evidence. Concretely, for CNN-based (or successor-architecture) attribution of paintings such as Frank and Frank's Rembrandt and van Gogh systems:

  1. (Interpretability of the artist-hand cues) Develop and validate tools that characterize what visual evidence the trained classifier actually uses to distinguish an artist's hand from that of imitators, going beyond localization tools (saliency or occlusion maps that show where the network looks) to identify the painterly cues themselves, and demonstrate that the characterization is causal rather than post-hoc narrative.

  2. (Cause disambiguation within a painting) Design controlled experiments and methods that separate, within a single painting, the signatures of the master from those of any other contributor and resolve each region-level 'not-by-the-master' classification to a single determinate cause — a workshop assistant's or pupil's hand, a restorer's intervention, a forger's imitation, retouching of the digitized image, or the master's own atypical manner — rather than leaving an unresolved coexistence of possibilities.

  3. (Cross-genre reliability) Determine, with controlled evaluation, the extent to which a classifier trained on one genre of an artist's work (e.g., van Gogh landscapes) transfers to other genres of the same artist (e.g., still lifes, portraits), covering more than one untrained genre, replacing the source's single caveat-laden test with a systematic, per-genre account.

These three components jointly serve one objective stated by the source: 'the ability to identify the different authors of a work of art, or at least to distinguish between the hand of the master and the work of others'. All three are required and none is independently sufficient: the problem is resolved only when each component has been addressed. The account sought is subject to the epistemic bound the source identifies: because an artist's full stylistic range and the limitless ways it can be imitated always exceed any training distribution, generalization claims are qualified, not absolute. The required evidence must ultimately be resolvable against curated, well-documented corpora and real paintings with independently established (scholarly-consensus or conservation-documented) ground truth, in combination with traditional scholarship and scientific analysis rather than as a replacement for them.

References

  1. Steven J. Frank and Andrea M. Frank, 'Analysis of Dutch Master Paintings with Convolutional Neural Networks', arXiv:2002.05107 (2020). DOI: 10.48550/arXiv.2002.05107, https://arxiv.org/abs/2002.05107 (earlier arXiv version carried the title 'Rembrandts and Robots: Using Neural Networks to Explore Authorship in Painting').
  2. J. Li, L. Yao, E. Hendriks and J. Z. Wang, 'Rhythmic Brushstrokes Distinguish van Gogh from His Contemporaries: Findings via Automated Brushstroke Extraction', IEEE Transactions on Pattern Analysis and Machine Intelligence 34(6):1159–1176 (2012). DOI: 10.1109/TPAMI.2011.203, https://doi.org/10.1109/TPAMI.2011.203
  3. Nanne van Noord, Ella Hendriks and Eric Postma, 'Toward Discovery of the Artist's Style: Learning to recognize artists by their artworks', IEEE Signal Processing Magazine 32(4):46–54 (2015). DOI: 10.1109/MSP.2015.2406955, https://doi.org/10.1109/MSP.2015.2406955
  4. F. Ji, M. S. McMaster, S. Schwab, G. Singh, L. N. Smith, S. Adhikari, M. O'Dwyer, F. Sayed, A. Ingrisano, D. Yoder, E. S. Bolman, I. T. Martin, M. Hinczewski and K. D. Singer, 'Discerning the painter's hand: machine learning on surface topography', Heritage Science 9:152 (2021). DOI: 10.1186/s40494-021-00618-w, https://doi.org/10.1186/s40494-021-00618-w
  5. Yoon-Seok Choi, Jae Woo Kim, Soonchul Jung, Hyeong-Ju Jeon and Jin Seo Kim, 'Efficient Analytical Data Processing and Visualization Methods for Supporting Scientific Connoisseurship on Oil Paintings', 2022 13th International Conference on Information and Communication Technology Convergence (ICTC) (2022). DOI: 10.1109/ICTC55196.2022.9952845, https://doi.org/10.1109/ICTC55196.2022.9952845
  6. Steven J. Frank and Andrea M. Frank, 'Analysis, attribution, and authentication of drawings with convolutional neural networks', International Journal of Arts and Technology 14(3) (2022). DOI: 10.1504/IJART.2022.128455, https://doi.org/10.1504/IJART.2022.128455 (preprint: DOI 10.21203/rs.3.rs-2520519/v1)
  7. Ludovica Schaerf, Carina Popovici and Eric Postma, 'Art Authentication with Vision Transformers', arXiv:2307.03039 (2023). DOI: 10.48550/arXiv.2307.03039, https://arxiv.org/abs/2307.03039
  8. Hassan Ugail, David G. Stork, Howell Edwards, Steven C. Seward and Christopher Brooke, 'Deep transfer learning for visual analysis and attribution of paintings by Raphael', Heritage Science 11:268 (2023). DOI: 10.1186/s40494-023-01094-0, https://doi.org/10.1186/s40494-023-01094-0
  9. Cheshta Arora and Debarun Sarkar, 'On the Injunction of XAIxArt', arXiv:2309.06227 (2023). DOI: 10.48550/arXiv.2309.06227, https://arxiv.org/abs/2309.06227
  10. Ravidu Suien Rammuni Silva, Ahmad Lotfi, Isibor Kennedy Ihianle, Golnaz Shahtahmassebi and Jordan J. Bird, 'ArtBrain: An Explainable end-to-end Toolkit for Classification and Attribution of AI-Generated Art and Style', arXiv:2412.01512 (2024). DOI: 10.48550/arXiv.2412.01512, https://arxiv.org/abs/2412.01512
  11. Tarian Fu, Javier Conde, Gonzalo Martínez, Pedro Reviriego, Elena Merino-Gómez and Fernando Moral, 'Artificial Intelligence and Misinformation in Art: Can Vision Language Models Judge the Hand or the Machine Behind the Canvas?', arXiv:2508.01408 (2025). DOI: 10.48550/arXiv.2508.01408, https://arxiv.org/abs/2508.01408
  12. A. Afifi, A. Kalimullin, S. Korchagin and I. Kudryashov, 'A deep learning perspective on Rubens' attribution', arXiv:2511.22667 (2025). DOI: 10.48550/arXiv.2511.22667, https://arxiv.org/abs/2511.22667
  13. Nicola Fanelli, Pasquale De Marinis, Raffaele Scaringi, Eva Cetinic, Gennaro Vessio and Giovanna Castellano, 'Understanding How MLLMs Describe Artworks Using Token Activation Maps', arXiv:2606.27947 (2026). DOI: 10.48550/arXiv.2606.27947, https://arxiv.org/abs/2606.27947
  14. Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh and Dhruv Batra, 'Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization', arXiv:1610.02391 (2016). DOI: 10.48550/arXiv.1610.02391, https://arxiv.org/abs/1610.02391
  15. Ruifang Wang, 'Deep learning–driven authentication and forgery detection methods for oil paintings', Signal Processing: Image Communication 147:117591 (2026). DOI: 10.1016/j.image.2026.117591, https://doi.org/10.1016/j.image.2026.117591