Pith. sign in

REVIEW 3 major objections 5 minor 62 references

Interpretable Classification of Levantine Ceramic Thin Sections via Neural Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Transfer learning lifts ceramic thin-section classification to 92% accuracy.

desk verdict Useful new dataset and a thoughtful XAI study, but the headline accuracies do not measure generalization to new ceramics because the random image-level split leaks samples across train and test. read the letter →

arxiv 2506.12250 v1 pith:OSJX7ATM submitted 2025-06-13 cs.CE

Sara Capriotti , Alessio Devoto , Simone Scardapane , Silvano Mignardi , Laura Medeghini This is my paper · ORCID
classification cs.CE
keywords ceramicthinsectionspetrographicfabricsLevantinearchaeologytransferlearningResNet18VisionTransformerGuidedGrad-CAMattentionmaps
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that deep neural networks can serve as a practical, transparent aid for classifying ancient ceramic thin sections by petrographic fabric, the mineral recipe used to make pottery. On a dataset of 1,424 polarized-light images from 178 Bronze and Iron Age Levantine samples grouped into ten fabrics, it reports that transfer learning from a large natural-image pre-training pushes a ResNet18 convolutional network to 92.11% accuracy and a Vision Transformer to 88.34%, compared with 76.33% and 61.01% when the same architectures are trained from scratch. It further claims that Guided Grad-CAM and attention maps reveal the models concentrating on the same diagnostic inclusions, calcite, quartz, dolomite, shell, and basalt, that human petrographers use. If these claims hold, archaeometric studies gain a fast, reproducible, and interpretable first-pass classification tool.

What carries the argument

The load-bearing machinery is transfer learning applied to two standard image-classification architectures, ResNet18 and a Vision Transformer, and then opened up by two explainability tools. ResNet18 is an 18-layer convolutional network whose skip connections let it train deeper feature hierarchies; the ViT here divides each 224 by 224 image into fixed 16 by 16 patches and processes them with 12 transformer blocks of multi-head self-attention, which lets it relate mineral inclusions across the whole section. Both are initialized with pre-trained natural-image weights, and only the final classification layers are retrained on the 1,424 thin-section images. Guided Grad-CAM then produces gradient-weighted activation maps from the CNN's last convolutional layer, and attention maps show which patches the ViT's heads weight most; together they turn a classification score into a visual claim about which minerals drove the decision.

What would settle it

Take the same 178 samples and split at the sample level, training on images from some samples and testing only on images from samples never used in training, then compare accuracy with the reported 92.11% and 88.34%; if accuracy drops substantially, the image-level split was inflating performance and the models have not learned to generalize to new ceramics.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that transfer learning solves the small-data problem in ceramic petrography: pre-trained models classify thin-section images into their known fabrics with high accuracy, while from-scratch models do not. The authors report ResNet18 at 92.11% accuracy and ViT at 88.34%, with precision, recall, and F1 all above 86% for the pre-trained ViT and above 90% for the pre-trained ResNet18. They also find that the errors are not random: both models confuse carbonate fabric A with carbonate fabric B, and one sample that mixes microfossils and dispersed quartz sits in exactly the ambiguous zone that human experts would struggle with. The explainability evidence is the second half of the claim: Guided Grad-CAM heatmaps track calcite crystals through rotation even into extinction, selectively highlight basalt fragments in the basalt fabric, and ignore stray carbonate fragments in a quartz fabric, while ViT attention maps converge on dolomite crystals and shell fragments in deeper layers. That convergence is what the authors present as evidence that the models learned mineralogy, not image artifacts.

Load-bearing premise

The evaluation assumes that a random 80:20 split at the level of individual images measures generalization to new ceramics, even though eight images come from each sample and images of the same sample can appear in both training and test sets, so the reported accuracy does not establish performance on samples the model has never seen.

Editorial extensions

If this is right

  • A first-pass screen: archaeologists can use the trained models to sort large ceramic assemblages into candidate petrographic fabrics before spending expert time on manual analysis.
  • Architecture-independence: because a transformer matches a CNN once both are pre-trained, the method does not depend on choosing a particular network family; transfer learning is the decisive ingredient.
  • Expert-aligned features: the models' focus on the same inclusions that define fabrics means automated scores can be reported alongside the mineral evidence supporting them, which helps archaeologists trust and audit the output.
  • Error-model insight: the systematic confusion between carbonate fabrics A and B identifies genuinely intermediate samples and could prioritize them for closer study rather than treating them as random failures.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A sample-level split would almost certainly produce lower accuracy than the reported 92.11% and 88.34%, because the current random 80:20 image split can place images of the same physical sample on both sides; the paper itself lists testing on unseen sections as future work.
  • The saliency maps' agreement with mineralogy suggests the same pipeline could flag candidate diagnostic minerals for poorly understood fabrics, effectively using fine-tuned vision models as hypothesis generators in archaeometry.
  • The method should transfer to other imaging-based materials identification tasks, such as sourcing stone or metal inclusions, wherever a labeled reference set of micrographs exists.
  • Pairing the image models with geochemical measurements, which the paper mentions as a future direction, might break the carbonate A/B ambiguity by adding compositional information that the textures alone cannot separate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript applies transfer learning with ResNet18 and a Vision Transformer (described in Appendix A as DeiT-Small) to classify 1,424 thin-section images from 178 Levantine ceramic samples into ten petrographic fabrics. The pre-trained ResNet18 reaches 92.11% accuracy and the pre-trained ViT 88.34% on a held-out test set, with from-scratch models performing substantially worse. Guided Grad-CAM and attention maps are used to argue that the models focus on mineralogical features such as calcite, quartz, dolomite, and basalt inclusions. The abstract and conclusion present these results as evidence that deep learning can support ceramic fabric classification in archaeometric studies.

Significance. If the reported accuracies were valid estimates of generalization to new ceramic samples, this would be a useful contribution: it appears to be the first application of ViTs to ceramic fabric classification, includes a transfer-learning ablation, and provides an interpretability analysis with rotation sequences and both PPL/XPL images. The curated dataset of 178 samples from six Levantine sites is also of archaeological value. However, the evaluation protocol is the load-bearing element of the paper's central claim, and as described below it does not support the generalizability statement in the abstract and conclusion.

major comments (3)
  1. [Section 2, Table 2] The evaluation split is performed at the image level, not the sample level: the text states that the 1,424 images are 'randomly split into training and testing sets with a ratio of 80:20' with no grouping by ceramic sample. Because each of the 178 samples contributes eight images, the probability that a given sample has images in both partitions is approximately 1 - 0.8^8 ≈ 0.83, so roughly 30 or more samples are represented in both training and test sets. The 92.11% and 88.34% accuracies in Table 2 therefore measure recognition of near-duplicate views of thin sections seen during training, not classification of new ceramics. A sample-level split (grouping all eight images of a sample into one partition) or leave-one-sample-out cross-validation is required for the abstract's and Section 5's generalization claims to be supported.
  2. [Section 3, Appendix A] The mean and standard deviation in Table 2 are computed over three random initializations, but the text describes a single random 80:20 split of the data. The reported uncertainty therefore ignores the variability due to the data split, which is typically the dominant source of variance for a dataset of this size. The authors should either report results over multiple sample-level splits or use repeated cross-validation so that the reported confidence intervals reflect the actual uncertainty in the accuracy estimates.
  3. [Section 5] The manuscript itself acknowledges in Section 5 that 'it will be essential to evaluate the model on additional sets of thin-section images, distinct from those used during training' and lists this as a future step. This is not merely a future limitation; it directly contradicts the present-tense generalization implied by the abstract ('a ResNet18 model achieving 92.11% accuracy') and by the conclusion ('yielded high classification accuracy'). If the experiments cannot be rerun with a proper sample-level split, the abstract and conclusion should be reworded to state explicitly that the reported accuracies are on test images that may come from the same samples as training images.
minor comments (5)
  1. [Section 2] The text says 'the inclusion of Elba provides important comparative data'; the site name is spelled 'Ebla' throughout the rest of the paper, including Table 1.
  2. [Section 2] There is a typo: 'primarly dating to the Bronze Age' should read 'primarily dating to the Bronze Age'.
  3. [Appendix A] The main text refers to the transformer model as 'ViT' and describes a 12-block, 384-hidden-dimension architecture, while Appendix A identifies it as 'DeiT-Small' with 6 heads. Please clarify whether the model is a DeiT-Small or a custom ViT, since DeiT-Small has specific architectural and training details that differ from the base ViT.
  4. [Figure 3] The caption says 'accuracy trends' without specifying that the testing accuracy was monitored only for visualization and not used for model selection; the main text does clarify this, but the caption could be more explicit to avoid confusion.
  5. [General] No data availability or code availability statement is provided. Given the claim of a 'reproducible and efficient methodology', please include a statement on where the dataset and code can be accessed, or explain any restrictions.

Circularity Check

1 steps flagged · score 6.0 of 10

Image-level 80:20 split places images of the same ceramic samples in both train and test, so the reported accuracies measure within-sample view recognition rather than classification of new ceramics.

  1. fitted input called prediction [Section 2, Materials and Methods (image acquisition and dataset split paragraphs); Section 5 Conclusion]
    "For each ceramic section, we took multiple images under plane-polarized light (PPL) and cross-polarized light (XPL) at magnifications of 2.5X and 10X, resulting in a total of eight images per section ... The final dataset, consisting of 1,424 images grouped in 10 classes, is randomly split into training and testing sets with a ratio of 80:20."

    The random split is performed on individual images, not on the 178 ceramic samples. Because every sample contributes eight images, the training and testing partitions almost certainly contain images from the same sample. The model is therefore fitted on some views of a given thin section and then 'predicts' the fabric label of a different view of that same section. The reported test accuracies (92.11% ResNet18, 88.34% ViT) thus reduce, by construction, to recognition of near-duplicate views of samples already seen during training, rather than to generalization to previously unseen ceramics. The paper itself concedes in Section 5 that evaluating on thin-section images 'distinct from those used during training' is a future step, confirming that the current test set is not sample-disjoint.

full rationale

The only load-bearing circular step is the evaluation split. The paper randomly splits 1,424 images, eight per ceramic sample, into an 80:20 train/test partition without grouping by sample, so test images are drawn from the same thin sections seen in training. The headline accuracies are therefore inflated estimates of within-sample view recognition, not of classification of new ceramics, and the paper explicitly acknowledges in Section 5 that evaluation on distinct thin-section images remains future work. Apart from this evaluation leakage, the study is not derivationally circular: the petrographic-fabric labels are external classifications from prior papers, no fitted constants are relabeled as predictions, and the explainability analysis is an empirical check rather than a self-referential derivation. The pretrained-versus-scratch transfer-learning comparison remains informative even with the leaked split, since both models are evaluated under the same protocol.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The central claim rests on two external inputs: the correctness of the fabric labels from prior petrographic studies (which the authors also published) and the validity of the image-level split as an evaluation of generalization. No new physical entities are introduced, and no constants are fitted to make a derivation work.

assumptions (3)
  • domain assumption The 10 petrographic fabric labels assigned in references [48-55] are correct and consistent.
    The paper uses these labels as ground truth without re-verification.
  • ad hoc to paper Images from the same ceramic sample are independent enough that splitting at the image level is a valid evaluation.
    This assumption is not stated but is implicit in the random image-level split; it is questionable because samples yield eight correlated images.
  • domain assumption Transfer learning from ImageNet provides useful features for petrographic micrographs.
    The paper relies on this to achieve high accuracy with limited data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interpretable Classification of Levantine Ceramic Thin Sections via Neural Networks." pith.science (2026). https://pith.science/paper/OSJX7ATM

@misc{pith2026250612250,
  author       = {Pith},
  title        = {Pith review of: Interpretable Classification of Levantine Ceramic Thin Sections via Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OSJX7ATM}},
  note         = {Machine review of arXiv:2506.12250}
}
read the original abstract

Classification of ceramic thin sections is fundamental for understanding ancient pottery production techniques, provenance, and trade networks. Although effective, traditional petrographic analysis is time-consuming. This study explores the application of deep learning models, specifically Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), as complementary tools to support the classification of Levantine ceramics based on their petrographic fabrics. A dataset of 1,424 thin section images from 178 ceramic samples belonging to several archaeological sites across the Levantine area, mostly from the Bronze Age, with few samples dating to the Iron Age, was used to train and evaluate these models. The results demonstrate that transfer learning significantly improves classification performance, with a ResNet18 model achieving 92.11% accuracy and a ViT reaching 88.34%. Explainability techniques, including Guided Grad-CAM and attention maps, were applied to interpret and visualize the models' decisions, revealing that both CNNs and ViTs successfully focus on key mineralogical features for the classification of the samples into their respective petrographic fabrics. These findings highlight the potential of explainable AI in archaeometric studies, providing a reproducible and efficient methodology for ceramic analysis while maintaining transparency in model decision-making.

Figures

Figures reproduced from arXiv: 2506.12250 by the authors.

Figure 1
Figure 1. Map of the Levant showing the selected archaeological sites. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Thin section images of representative samples of the ten petrographic [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Classification performance of ResNet18 and Vision Transformer (ViT). (a) [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Comparison of ResNet18 and ViT misclassified predictions for class 2, [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Guided Grad-CAM applied to pre-trained ResNet18. The figure shows the [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Guided Grad-CAM applied to the rotation sequence of sample TF4 from [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Guided Grad-CAM applied to the rotation sequence of sample TFMcamp1 [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Attention maps of representative samples from carbonate A (sample [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 59 canonical work pages

  1. [1]

    Steiner and A

    M. Steiner and A. E. Killebrew. The Oxford Handbook of the Archaeology of the Levant: c. 8000-332 BCE. Oxford Handbooks, 2014

    work page 2014

  2. [2]

    Greenberg

    R. Greenberg. The archaeology of the bronze age levant. Cambridge University Press, 2019

    work page 2019

  3. [3]

    Greenberg

    R. Greenberg. Traveling in (world) time: transformation, commoditization, and the beginnings of urbanism in the Southern Levant. Oxbow Books, 2011

    work page 2011

  4. [4]

    H. Weiss. The northern levant during the intermediate bronze age. In Margreet L. Steiner and Ann E. Killebrew, editors, The Oxford Handbook of the Archaeology of the Levant: c. 8000–332 BCE, pages 401–412. Oxford University Press, Oxford, 2013

    work page 2013

  5. [5]

    Urbanism

    M. S. Chesson and G. Philip. Tales of the City? “Urbanism” in the Early Bronze Age Levant from Mediterranean and Levantine Perspectives. Horizon, 1(2):26–37, 2003

    work page 2003

  6. [6]

    P. S. Quinn. Ceramic Petrography: The Interpretation of Archaeological Pottery & Related Artefacts in Thin Section. Archaeopress Publishing, 2013

    work page 2013

  7. [7]

    Orton and M

    C. Orton and M. Hughes. Pottery in Archaeology. Cambridge University Press, 2013

    work page 2013

  8. [8]

    G. Yu, Z. Jin, L. Chen, F. Wang, X. Wang, X Wu, A. Fan, and Q. Xia. Analyzing the earliest Chinese proto-porcelain: Study on the materials from Liaotianjianshan kiln sites, Dehua County, Fujian Province (China). Ceramics International, 44(17):21648–21655, 2018

    work page 2018

Show all 62 references
  1. [9]

    I. K. Whitbrad. The characterisation of argillaceous inclusions in ceramic thin sections. Archaeometry, 28(1):79–88, 1986

    1986

  2. [10]

    I. K. Whitebread. Greek transport amphorae. A petrological and archaeological study. British School of Athens., 1995

    1995

  3. [11]

    I. K. Whitbread and A. Hunt. Fabric description of archaeological ceramics. The Oxford handbook of archaeological ceramic analysis. Oxford University Press, 2017

    2017

  4. [12]

    Y. Goren. The Southern Levant in the Early Bronze Age IV: The Petrographic Perspective. Bulletin of the American Schools of Oriental Research , 303(1):33–72, 1996

    1996

  5. [13]

    Greenberg and N

    R. Greenberg and N. Porat. A Third Millennium Levantine Pottery Production Center: Typology, Petrography, and Provenance of the Metallic Ware of Northern Israel and Adjacent Regions. Bulletin of the American Schools of Oriental Research , 301(1):5–24, 1996

    1996

  6. [14]

    Y. Goren. Ceramic Technology and Provenance at Khirbat Iskandar. In S. Richard, J. C. Long, P. S. Holdorf, and G. Peterman, editors, Archaeological Expedition to Khirbat Iskandar and its Environs, Jordan: Khirbat Iskandar Final Report on the Early Bronze IV Area C ’Gateway’ an...

    2010

  7. [15]

    D’Andrea

    M. D’Andrea. The Southern Levant in Early Bronze IV: issues and perspectives in the pottery evidence. Sapienza Univ. di Roma, Dipt. di Scienze dell’Antichit` a 20XX, 2015

    2015

  8. [16]

    Maritan, C

    L. Maritan, C. Mazzoli, V. Michielin, D. Morandi Bonacossi, M. Luciani, and G. Molin. The provenance and production technology of Bronze Age and Iron Age pottery from Tell Mishrifeh/Qatna (Syria). Archaeometry, 47(4):723–744, 2005

    2005

  9. [17]

    Badreshany, G

    K. Badreshany, G. Philip, and M. Kennedy. The development of integrated regional economies in the Early Bronze Age Levant: new evidence from ‘Combed Ware’ jars. Levant, 42(1-2):160–196, 2020

    2020

  10. [18]

    Tumolo and K

    V. Tumolo and K. Badreshany. The ‘Combed Ware’ storage and transport vessels from Khirbet ez-Zeraqon: a reappraisal of the EB II-III evidence in light of recent studies. Journal of Ancient Egyptian Interconnections, 37:301–324, 2023

    2023

  11. [19]

    princely

    G. Caspari and P. Crespo. Convolutional neural networks for archaeological site detection–finding “princely” tombs. Journal of Archaeological Science, 110, 2019

    2019

  12. [20]

    Guyot, M

    A. Guyot, M. Lennon, T. Lorho, and L. Hubert-Moy. Combined Detection and Segmentation of Archeological Structures from LiDAR Data Using a Deep Learning Approach. Journal of Computer Applications in Archaeology, 4(1):1, 2021

    2021

  13. [21]

    Ø. D. Trier, D. C. Cowley, and A. U. Waldeland. Using deep neural networks on airborne laser scanning data: Results from a case study of semi-automatic mapping of archaeological Interpretable Classification of Levantine Ceramic Thin Sections via Neural Networks 21 topography o...

    2019

  14. [22]

    Troiano, E

    M. Troiano, E. Nobile, F. Mangini, M. Mastrogiuseppe, C. Conati Barbaro, and F. Frezza. A comparative analysis of the bayesian regularization and levenberg–marquardt training algorithms in neural networks for small datasets: A metrics prediction of neolithic laminar artefacts....

    2024

  15. [23]

    Nobile, M

    E. Nobile, M. Troiano, F. Mangini, M. Mastrogiuseppe, J. Verdi, F. Frezza, C. Conati Barbaro, and A. Gopher. Neural network analysis for predicting metrics of fragmented laminar artifacts: a case study from MPPNB sites in the Southern Levant. Scientific Reports, 14, 2024

    2024

  16. [24]

    Siozos, N

    P. Siozos, N. Hausmann, M. Holst, and D. Anglos. Application of laser-induced breakdown spectroscopy and neural networks on archaeological human bones for the discrimination of distinct individuals. Journal of Archaeological Science: Reports , 35:102769, 2021

    2021

  17. [25]

    Zhang, I

    J. Zhang, I. Fang, H. Wu, A. Kaushik, A. Rodriguez, H. Zhao, J. Zhang, Z. Zheng, R. Iovita, and C. Feng. Luwa dataset: Learning lithic use-wear analysis on microscopic images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563–22573, 2024

    2024

  18. [26]

    S. H. Bickler. Machine Learning Arrives in Archaeology. Advances in Archaeological Practice , 9(2):186–191, 2021

    2021

  19. [27]

    L. M. Pawlowicz and C. E. Downum. Applications of deep learning to decorated ceramic typology and classification: A case study using Tusayan White Ware from Northeast Arizona. Journal of Archaeological Science, 130, 2021

    2021

  20. [28]

    H¨ orr, E

    C. H¨ orr, E. Lindinger, and Brunnett G. Machine learning based typology development in archaeology. Journal on Computing and Cultural Heritage , 7(1):1–23, 2014

    2014

  21. [29]

    Navarro, C

    P. Navarro, C. Cintas, M. Lucena, J. M. Fuertes, C. Delrieux, and M. Molinos. Learning feature representation of Iberian ceramics with automatic classification models. Journal of Cultural Heritage, 48:65–73, 2021

    2021

  22. [30]

    Y. Qi, M. Z. Qiu, H. Z. Jing, Z. Q. Wang, C. L. Yu, J. F. Zhu, F. Wang, and T. Wang. End-to-end ancient ceramic classification toolkit based on deep learning: A case study of black glazed wares of Jian kilns (Song Dynasty, Fujian province). Ceramics International , 48(23):3451...

    2022

  23. [31]

    Ruschioni, D

    G. Ruschioni, D. Malchiodi, A. M. Zanaboni, and L. Bonizzoni. Supervised learning algorithms as a tool for archaeology: Classification of ceramic samples described by chemical element concentrations. ournal of Archaeological Science, 49, 2023

    2023

  24. [32]

    Anglisano, L

    A. Anglisano, L. Casas, I. Queralt, and R. Di Febo. Supervised Machine Learning Algorithms to Predict Provenance of Archaeological Pottery Fragments. Sustainability, 14, 2022

    2022

  25. [33]

    Ramil, A.J

    A. Ramil, A.J. L´ opez, and A. Y´ a˜ nez. Application of artificial neural networks for the rapid classification of archaeological ceramics by means of laser induced breakdown spectroscopy (libs). Applied Physics A , pages 197–202, 2008

    2008

  26. [34]

    Aprile, G

    A. Aprile, G. Castellano, and G. Eramo. Combining image analysis and modular neural networks for classification of mineral inclusions and pores in archaeological potsherds. Journal of Archaeological Science, 50:262–272, 2014

    2014

  27. [35]

    M. Lyons. Ceramic Fabric Classification of Petrographic Thin Sections with Deep Learning. Journal of computer applications in archaeology , 4(1):188–201, 2021

    2021

  28. [36]

    Lyons, F

    M. Lyons, F. Fecher, and M. Reindel. From LiDAR to deep learning: A case study of computer- assisted approaches to the archaeology of Guadalupe and northeast Honduras. it - Information Technology, 64(6):233–246, 2022

    2022

  29. [37]

    LeCun, Y

    Y. LeCun, Y. Bengio, and G. Hinton. Deep Learning. Nature, 521(7553):436–444, 2015

    2015

  30. [38]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, Zhai X., T. Unterthiner, M. Dehghani, M Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. arXiv preprint arXiv:2010.11929 , 2020

    2010 arXiv

  31. [39]

    Raghu, T

    M. Raghu, T. Unterthiner, S. Kornblith, C. Zhang, and A. Dosovitskiy. Do Vision Transformers See Like Convolutional Neural Networks?. Advances in neural information processing systems , Interpretable Classification of Levantine Ceramic Thin Sections via Neural Networks 22 34:1...

    2021

  32. [40]

    LiHua, Z

    Y. LiHua, Z. WenBo, and Q. WangRen. Chronological classification of ming and qing dynasty ceramics images based on an enhanced resnet50 model. STAR: Science & Technology of Archaeological Research, 11(1):e2498260, 2025

    2025

  33. [41]

    Castelvecchi

    D. Castelvecchi. Can we open the black box of AI? Nature, 538(7623):20–23, 2016

    2016

  34. [42]

    Zhong, B

    X. Zhong, B. Gallagher, S. Liu, B. Kailkhura, A. Hiszpanski, and T. Y. J. Han. Explainable machine learning in materials science. Computational Materials , 8, 2022

    2022

  35. [43]

    A. B. Arrieta, N. D ´ ıaz-Rodr ´ ıguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garcia, S. Gil Lopez, D. Molina, R. Benjamins, R. Chaila, and F. Herrera. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible ...

    2020

  36. [44]

    Oviedo, J

    F. Oviedo, J. L. Ferres, T. Buonassisi, and K. T. Butler. Interpretable and Explainable Machine Learning for Materials Science and Chemistry. Accounts of Materials Research , 3(6):597–607, 2022

    2022

  37. [45]

    Montavon, W

    G. Montavon, W. Samek, and K. R. M¨ uller. Methods for interpreting and understanding deep neural networks. Digital Signal Processing, 73:1–15, 2018

    2018

  38. [46]

    Genovese, A

    D. Genovese, A. Sgroi, A. Devoto, S. Valentine, L. Wood, C. Sebastiani, S. Giagu, M. D’Onofrio, and S. Scardapane. Mixture-of-Experts Graph Transformers for Interpretable Particle Collision Detection., 2025

    2025

  39. [47]

    El-Hajj, O

    H. El-Hajj, O. Eberle, A. Merklein, A. Siebold, N. Shlomi, J. B¨ uttner, J. Martinetz, K. R. M¨ uller, G. Montavon, and M. Valleriani. Explainability and transparency in the realm of digital humanities: toward a historian XAI. International Journal of Digital Humanities , 5(2)...

    2023

  40. [48]

    Medeghini, M

    L. Medeghini, M. Sala, C. De Vito, and S. Mignardi. A forgotten centre of ceramic production in Southern Levant: Preliminary analytical study of the Early Bronze Age pottery from Tell el-Far‘ah North (West Bank). Ceramics International, 45(9):11457–11467, 2019

    2019

  41. [49]

    Medeghini, S

    L. Medeghini, S. Mignardi, C. De Vito, N. Macro, M. D’Andrea, and S. Richard. New insights on Early Bronze Age IV pottery production and consumption in the southern Levant: The case of Khirbat Iskandar, Jordan. Ceramics International, 42(16):18991–19005, 2016

    2016

  42. [50]

    Botticelli, S

    M. Botticelli, S. Mignardi, C. De Vito, Y. W. Liao, D. Montanari, M. Shakarna, L. Nigro, and L. Medeghini. Variability in pottery production at Khalet al-Jam’a necropolis, Bethlehem (West Bank): From the Early-Middle Bronze to the Iron Age. Ceramics International, 46(10):16405...

    2020

  43. [51]

    Botticelli, S

    M. Botticelli, S. Mignardi, C. De Vito, M. Sala, G. Mazzotta, L. Da Silva Gondim, and L. Medeghini. The ‘metallic ware’ from Tell el Far’ah North (West Bank): Petrography, technology, and provenance of a hidden ceramic industry. Ceramics International, 48(1):1366– 1374, 2022

    2022

  44. [52]

    M. Sala. The EB II ‘metallic ware’ from Tell el-Far ‘ah North (West Bank): typology, technology and petrography of a ceramic industry of the central hill country. Levant, 55(2):146–171, 2023

    2023

  45. [53]

    Palace of the Copper Axes

    L. Medeghini, L. Fabrizi, S. De Vito, C. Mignardi, L. Nigro, and C. Gallo, E. Fiaccavento. The ceramic of the “Palace of the Copper Axes” (Khirbet al-Batrawy, Jordan): A palatial special production. Ceramics International, 42(5):5952–5962, 2016

    2016

  46. [54]

    Ballirano, C

    P. Ballirano, C. De Vito, L. Medeghini, S. Mignardi, V. Ferrini, P. Matthiae, D. Bersani, and P. Lottici. A combined use of optical microscopy, X-ray powder diffraction and micro- Raman spectroscopy for the characterization of ancient ceramic from Ebla (Syria). Ceramics Intern...

    2014

  47. [55]

    Medeghini and L

    L. Medeghini and L. Nigro. Khirbet al-Batrawy ceramics: a systematic mineralogical and petrographic study for investigating the material culture. Periodico di Mineralogia, 86(1), 2017

    2017

  48. [56]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 770–778., 2016

    2016

  49. [57]

    J. Deng, W. Dong, R. Socher, L. J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical Interpretable Classification of Levantine Ceramic Thin Sections via Neural Networks 23 image database. IEEE conference on computer vision and pattern recognition , pages 248–255., 2009

    2009

  50. [58]

    D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimization., 2014

    2014

  51. [59]

    S. Yang, W. Xiao, M. Zhang, S. Guo, J. Zhao, and F. Shen. Image Data Augmentation for Deep Learning: A Survey., 2023

    2023

  52. [60]

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. International Journal of Computer Vision , 128(2):336–359, October 2019

    2019

  53. [61]

    Kashefi, L

    R. Kashefi, L. Barekatain, M. Sabokrou, and F. Aghaeipoor. Explainability of Vision Transformers: A Comprehensive Review and New Perspectives. arXiv preprint arXiv:2311.06786 , 2023

    2023 arXiv

  54. [62]

    L. Gao, T. D. la Tour, H. Tillman, G. Goh, R. Troll, A. Radford, I. Sutskever, J. Leike, and J. Wu. Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093 , 2024

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.