Pith. sign in

REVIEW 3 major objections 4 minor 81 references

Vision transformers match human texture perception better than CNNs, and architecture—not training objective—seems to drive how textures are coded.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-10 09:25 UTC pith:NR2J5WBG

load-bearing objection Clean empirical package showing three ViTs track human texture odd-one-out better than VGG-19; the architecture claim is real but rests on a single CNN that also defines one stimulus class. the 3 major comments →

arxiv 2607.08321 v1 pith:NR2J5WBG submitted 2026-07-09 cs.CV

Texture Representations in Deep Vision Models: Comparing CNNs, Vision Transformers, and Human Perception

Ludovica de Paolis , Marco Baroni , Alessandro Laio , Eugenio Piasini This is my paper
classification cs.CV
keywords texture perceptionvision transformersconvolutional neural networkshuman psychophysicsinformation imbalancerepresentational alignmentodd-one-outarchitecture
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Standard models of biological vision lean on convolutional networks trained for object recognition, but it is unclear whether that alignment holds for texture perception—a core visual ability that is not the same as naming objects. This paper builds a continuum of textures from the same source images using three synthesis methods of increasing complexity, plus natural textures and objects, then measures how a CNN and three vision transformers represent those stimuli with a rank-based information measure. It also collects human odd-one-out judgments on the same materials. The transformers agree with one another and form stable, overlapping representations across texture complexities; the CNN does not. Human discrimination performance tracks the transformers far better than the CNN. The claim is that attention-based architectures may be better models of how humans process texture, and that network architecture, more than task or training signal, shapes texture representations.

Core claim

Texture representations align across three vision transformers but not between those transformers and a CNN; the transformers form similar representations for textures of different complexity, and human odd-one-out performance on textures is better predicted from transformer representations than from CNN representations. Architecture (attention versus convolution) is the factor the authors identify as driving the difference.

What carries the argument

Information Imbalance, a rank-based, asymmetric measure of shared local neighborhood structure between high-dimensional representation spaces; used both to compare models on the same stimulus set and to relate model geometry to human accuracy on texture pairs.

Load-bearing premise

That one convolutional network is representative enough of the whole CNN family that the gap with transformers can be blamed on architecture rather than on that particular model or on the fact that one texture type is defined by that same network.

What would settle it

Repeat the same Information Imbalance and human-correlation analyses with several other CNNs (and preferably transformers not trained only on ImageNet); if other CNNs match human texture judgments as well as the transformers, the architecture claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper compares texture representations in one CNN (VGG-19) and three Vision Transformers (CLIP, DINO-v2, iGPT) against human odd-one-out psychophysics. Stimuli form a complexity continuum (Noise, Victor–Conte, Portilla–Simoncelli, Gatys, DTD, ImageNet objects) generated from a shared DTD source. Using symmetrized Information Imbalance (II) on layer activations, the authors report that the three ViTs form mutually aligned representations of textures of varying complexity, that VGG-19 does not align with them except on high-semantic stimuli, and that human pairwise discrimination accuracy correlates strongly with ViT II scores (Pearson r ≈ 0.94–0.96) but not with VGG-19. They conclude that architecture (attention vs. convolution) is a primary driver of texture coding and that ViTs may better model human texture perception than CNNs.

Significance. If the architecture claim holds, the work usefully challenges the near-exclusive reliance on CNNs as models of biological vision for non-object tasks and supplies a concrete, rank-based pipeline (II + odd-one-out) that can be reused. Strengths include an independent human benchmark, a previously published metric (II), off-the-shelf models, and a carefully constructed continuum of textures that share source images. The human–ViT alignment result is the most novel and potentially impactful finding for computational vision science.

major comments (3)
  1. The central architecture claim (attention vs. convolution drives texture coding more than training objective) rests on a single CNN (VGG-19; Methods 2.2, Table 1). The Limitations section itself notes that other CNNs were omitted for compute reasons and only conjectures they would behave like VGG-19. Without at least one additional CNN family (e.g., ResNet-50 or AlexNet) run on the identical pipeline, the ViT–CNN gap cannot be securely attributed to architecture rather than to idiosyncrasies of VGG-19.
  2. Gatys (G) textures are defined by matching Gram matrices of VGG-19 feature maps (Methods 2.1, Appendix 6.3.3). When the same network is later asked to represent those textures, part of its internal geometry is tautological: the stimuli already live in its own second-order feature statistics. This can artificially inflate VGG–G similarity relative to other subsets and depress VGG–ViT alignment on the mid-to-high complexity textures that dominate the human correlation (Fig. 4b). The three ViTs never generated any of the stimuli, so their mutual consistency and human alignment are free of that circularity. A control that re-synthesizes G textures from a different backbone, or that excludes G from the human–model correlation, is needed before the architecture attribution can be trusted.
  3. Summary graphs (Figs. 2–3) and the human–model correlations (Fig. 4b) select the ‘best’ (most mutually predictive) layer pair for each model/dataset combination. This post-hoc selection can inflate apparent alignment. Reporting the full layer-wise matrices (already present in the appendix) as the primary result, or pre-registering a fixed relative-depth criterion, would make the claim more robust.
minor comments (4)
  1. Activation clipping to the 0.95 percentile is applied only to ViTs (Results §3); the choice of percentile and its effect on II should be justified or ablated.
  2. Figure 1 caption and Methods 2.1: the Object image is ‘chosen by hand for illustrative purposes’; a random or class-matched example would be preferable.
  3. Appendix Table 2 reports p-values for the human–model correlations; the main text should also state the number of pairs (n=6) so readers can judge degrees of freedom.
  4. Minor typographical issues: ‘neigborhoods’ (p. 3), ‘strenght’ (p. 6), inconsistent hyphenation of ‘odd-one-out’.

Circularity Check

1 steps flagged

No load-bearing circularity: central human–ViT alignments rest on independent psychophysics and a published metric; only mild self-reference via prior arXiv and VGG-defined Gatys stimuli.

specific steps
  1. other [Methods 2.1 / Appendix 6.3.3 + Limitations]
    "The algorithm by Gatys synthesizes naturalistic textures by computing local correlations among the feature maps extracted from a CNN pretrained on object recognition (VGG-19…). The algorithm’s backbone is VGG-19_bn… Due to limited computational resources, we could not include other CNN models in the analyses. If this could be done, based on the representational similarities for textures across different CNN architectures [de Paolis et al., 2026], we would expect them to show similar results as VGG-19"

    Gatys (G) stimuli are defined by matching VGG-19 second-order feature statistics; the same network is later used as the sole CNN whose texture representations are compared with ViTs and with humans. Combined with the self-citation that other CNNs would behave identically, this creates a mild non-independence for the architecture-attribution claim on the mid/high-complexity textures that dominate the human correlation. It does not, however, make any numerical II value or Pearson r equal to its input by construction.

full rationale

The paper is an empirical comparison, not a first-principles derivation. Features are extracted from four off-the-shelf pretrained models (VGG-19, CLIP, DINO-v2, iGPT), Information Imbalance (a previously published rank statistic of Glielmo et al.) is computed between layers/datasets, and the resulting pairwise II values are correlated with new human odd-one-out accuracies collected for this study. None of these steps reduces by construction to a fitted parameter, a definitional identity, or an unverified uniqueness claim. Self-citations to de Paolis et al. (2026) supply motivation (“texture quality does not correlate with object-recognition alignment”) and the public Gatys implementation; they are not required for the numerical results or the human correlations. The fact that Gatys textures are synthesized from VGG-19 Gram matrices while VGG-19 is also the sole CNN comparator is a potential validity confound for the architecture interpretation (and is openly noted in Limitations), but it does not force the reported II graphs or the Pearson correlations with human accuracy to equal their inputs. The three ViTs never generated any stimuli, their mutual consistency and human alignment are free of that loop, and the human data remain an external benchmark. Hence circularity is at most minor and non-load-bearing.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on the validity of Information Imbalance as a measure of representational similarity, on the assumption that the three synthesis algorithms produce a meaningful complexity continuum, on the representativeness of a single CNN, and on a handful of preprocessing choices (percentile clipping, best-layer selection, grayscale conversion). No new physical entities are postulated; free parameters are limited to analysis hyperparameters.

free parameters (4)
  • activation clipping percentile = 0.95
    ViT activations clipped at the 0.95 percentile to remove attention-sink outliers; choice affects the feature spaces fed to II.
  • best-layer selection for summary graphs
    Main figures report only the layer pair with largest 1−II; full matrices are in appendix but the narrative uses the selected values.
  • Gatys synthesis hyperparameters = 30000 steps
    Adam steps (30 000), layer weighting, and VGG-19_bn backbone fix the statistical content of the G textures.
  • Portilla–Simoncelli pyramid parameters = n=5,k=4,na=7,iter=10
    Depth 5, 4 orientations, 7-pixel autocorrelation, 10 iterations define the mid-complexity textures.
axioms (4)
  • domain assumption Information Imbalance (symmetrized) quantifies shared information between high-dimensional representational spaces via neighborhood ranks.
    Invoked throughout Results; taken from Glielmo et al. 2022 without re-derivation.
  • domain assumption The ordered set Noise < V&C < P&S < G < DTD constitutes a meaningful perceptual-complexity continuum for textures.
    Figure 1 and Methods 2.1; used to interpret both model and human graphs.
  • ad hoc to paper A single CNN (VGG-19) is sufficiently representative of the CNN family for the architecture comparison.
    Explicitly stated as a compute limitation; load-bearing for the claim that architecture, not model idiosyncrasy, drives the gap.
  • domain assumption Odd-one-out accuracy on 200 ms grayscale presentations indexes the same representational geometry that II measures in network layers.
    Section 3 correlation analysis; bridges psychophysics to internal features.

pith-pipeline@v1.1.0-grok45 · 21441 in / 3004 out tokens · 32690 ms · 2026-07-10T09:25:05.438748+00:00 · methodology

0 comments
read the original abstract

In computational vision science, Convolutional Neural Networks (CNNs) have emerged as a popular model of biological vision because of the alignment they can exhibit with neural and behavioral data in humans and animals. However, it remains unclear to what extent this alignment persists for visual tasks that extend beyond the canonical object recognition paradigm based on well defined semantic content. In this study, we diverge from the common object-centric view by focusing on another aspect of vision: texture perception. We consider textures of different complexity generated with three different algorithms from the same source images. Using a rank-based statistic, we quantify the information encoded in the internal representations of a CNN and three Vision Transformers (ViTs), and we compare the similarity of these representations to those inferred from human psychophysics data. We find that the representation of textures is aligned in different ViTs, but not between the ViTs and the CNN; that ViTs form similar representations for textures of different complexity; that human performance in recognizing textures can be better predicted from ViTs representations rather than CNN representations. Taken together, these results suggest that ViTs may capture more faithfully than CNNs how texture patterns are visually processed by humans, and that the representations of texture stimuli in computational models may be driven by the network architecture.

Figures

Figures reproduced from arXiv: 2607.08321 by Alessandro Laio, Eugenio Piasini, Ludovica de Paolis, Marco Baroni.

Figure 1
Figure 1. Figure 1: Example stimuli arranged along a putative spectrum of texture complexity. In the Noise sample, each pixel is sampled independently at random as black or white with equal probabil￾ity. The stimuli labeled as V&C, P&S, and G are obtained by the DTD image classified as “braided” (marked as “source”) and by application of the respective algorithm. The Object image was chosen by hand for illustrative purposes f… view at source ↗
Figure 2
Figure 2. Figure 2: Graphs representing the symmetrized mean II across models. For visualization purposes, [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Graphs representing the symmetrized mean II across datasets, excluding Object. For [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Human pairwise accuracy and its correlations with models performance. For illustrative [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Heatmaps for each model (rows) by each dataset (columns). For each matrix, the entries [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: (a) Classification: accuracy scores for a triplet of probes trained on early, middle, and [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Extended plot for Section 4.2, [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Extended plot for Section 4.3, [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Extended plot for Section 4.5, Figure 4a: This plot represents the relationships among texture subsets for each pair of odd image (see X axis) and and reference image (legend) as human accuracy scores (see Y axis) of all participants collected in the behavioral task. Each box summa￾rizes the distribution of mean accuracies for one combination of odd source and reference source. The scatter shows the indivi… view at source ↗
Figure 10
Figure 10. Figure 10: Example of the visual stimulus screen in the experimental design. The odd in this case is [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Illustration of the experimental flow, with time in ms. Each simulation is paced at 200ms, [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

81 extracted references · 81 canonical work pages · 5 internal anchors

  1. [1]

    Frontiers in Cellular Neuroscience , volume =

    Efficient Processing of Natural Scenes in Visual Cortex , author =. Frontiers in Cellular Neuroscience , volume =

    work page

  2. [2]

    Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil , date =. An

    work page

  3. [3]

    Proceedings of the Royal Society B: Biological Sciences , volume =

    Predisposed and Learned Preferences for Multipoint Visual Statistics in Visually Naive Newly Hatched Chicks , author =. Proceedings of the Royal Society B: Biological Sciences , volume =. doi:10.1098/rspb.2025.3157 , urldate =

    work page doi:10.1098/rspb.2025.3157 2025

  4. [4]

    doi:10.1073/pnas.91.18.8482 , url =

    Striate Cortex Extracts Higher-Order Spatial Correlations from Visual Textures , author =. doi:10.1073/pnas.91.18.8482 , url =. 8078907 , eprinttype =

    work page doi:10.1073/pnas.91.18.8482

  5. [5]

    Visual Processing of Informative Multipoint Correlations Arises Primarily in

    Yu, Yunguo and Schmid, Anita M and Victor, Jonathan D , date =. Visual Processing of Informative Multipoint Correlations Arises Primarily in. doi:10.7554/eLife.06604 , url =

    work page doi:10.7554/elife.06604

  6. [6]

    IRE Transactions on Information Theory , year =

    Julesz, Bela , title =. IRE Transactions on Information Theory , year =

    work page

  7. [7]

    and Julesz, B

    Caelli, T. and Julesz, B. , title =. Biological Cybernetics , year =

    work page

  8. [8]

    Nature , year =

    Julesz, Bela , title =. Nature , year =

    work page

  9. [9]

    Julesz, Bela and Gilbert, E. N. and Shepp, L. A. and Frisch, H. L. , title =. Perception , year =

    work page

  10. [10]

    Turner, M. R. , title =. Biological Cybernetics , year =

    work page

  11. [11]

    Journal of the Optical Society of America A , year =

    Malik, Jitendra and Perona, Pietro , title =. Journal of the Optical Society of America A , year =

    work page

  12. [12]

    , title =

    Portilla, Javier and Simoncelli, Eero P. , title =. International Journal of Computer Vision , year =

    work page

  13. [13]

    Zhu, S. C. and Liu, X. W. and Wu, Y. N. , title =. IEEE Transactions on Pattern Analysis and Machine Intelligence , year =

    work page

  14. [14]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li , title =. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    work page

  15. [15]

    Local Statistics in Natural Scenes Predict the Saliency of Synthetic Textures , journal =

    Tka. Local Statistics in Natural Scenes Predict the Saliency of Synthetic Textures , journal =. 2010 , volume =

    work page 2010

  16. [16]

    , title =

    Freeman, Jeremy and Simoncelli, Eero P. , title =. Nature Neuroscience , year =

    work page

  17. [17]

    and Conte, Mary M

    Victor, Jonathan D. and Conte, Mary M. , title =. Journal of the Optical Society of America A , year =

    work page

  18. [18]

    and Zoccolan, Davide and Rust, Nicole C

    DiCarlo, James J. and Zoccolan, Davide and Rust, Nicole C. , title =. Neuron , year =

    work page

  19. [19]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , year =

    Cimpoi, Mircea and Maji, Subhransu and Kokkinos, Iasonas and Mohamed, Sammy and Vedaldi, Andrea , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , year =

    work page

  20. [20]

    and Hong, Ha and Yamins, Daniel L

    Cadieu, Charles F. and Hong, Ha and Yamins, Daniel L. K. and Pinto, Nicolas and Ardila, Diego and Solomon, Ethan A. and Majaj, Najib and DiCarlo, James J. , title =. Proceedings of the National Academy of Sciences , year =

    work page

  21. [21]

    Deep Neural Networks Reveal a Gradient in the Complexity of Neural Representations across the Ventral Stream , journal =

    G. Deep Neural Networks Reveal a Gradient in the Complexity of Neural Representations across the Ventral Stream , journal =. 2015 , volume =

    work page 2015

  22. [22]

    and Victor, Jonathan D

    Yu, Yunguo and Schmid, Anita M. and Victor, Jonathan D. , title =. eLife , year =

    work page

  23. [23]

    and Kaiser,

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser,. Attention Is All You Need , booktitle =. 2017 , volume =

    work page 2017

  24. [24]

    and Ecker, Alexander S

    Gatys, Leon A. and Ecker, Alexander S. and Bethge, Matthias , title =. Advances in Neural Information Processing Systems , year =

    work page

  25. [25]

    and Ecker, Alexander S

    Gatys, Leon A. and Ecker, Alexander S. and Bethge, Matthias , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , year =

    work page

  26. [26]

    and Bashivan, Pouya and Kar, Kohitij and Schmidt, Kailyn and DiCarlo, James J

    Rajalingham, Rishi and Issa, Elias B. and Bashivan, Pouya and Kar, Kohitij and Schmidt, Kailyn and DiCarlo, James J. , title =. Journal of Neuroscience , year =

    work page

  27. [27]

    and Briguglio, John J

    Hermundstad, Ann M. and Briguglio, John J. and Conte, Mary M. and Victor, Jonathan D. and Balasubramanian, Vijay and Tka. Variance Predicts Salience in Central Sensory Processing , journal =. 2014 , volume =

    work page 2014

  28. [28]

    eLife , year =

    Caramellino, Riccardo and Piasini, Eugenio and Buccellato, Andrea and Carboncino, Anna and Balasubramanian, Vijay and Zoccolan, Davide , title =. eLife , year =

    work page

  29. [29]

    and Briguglio, John J

    Tesileanu, Tiberiu and Conte, Mary M. and Briguglio, John J. and Hermundstad, Ann M. and Victor, Jonathan D. and Balasubramanian, Vijay , title =. eLife , year =

    work page

  30. [30]

    and Conte, Mary M

    Victor, Jonathan D. and Conte, Mary M. and Chubb, Charles F. , title =. Annual Review of Vision Science , year =

    work page

  31. [31]

    and Heeger, David J

    Freeman, Jeremy and Ziemba, Corey M. and Heeger, David J. and Simoncelli, Eero P. and Movshon, J. Anthony , title =. Nature Neuroscience , year =

    work page

  32. [32]

    and Freeman, Jeremy and Movshon, J

    Ziemba, Corey M. and Freeman, Jeremy and Movshon, J. Anthony and Simoncelli, Eero P. , title =. Proceedings of the National Academy of Sciences , year =

    work page

  33. [33]

    Cerebral Cortex , year =

    Okazawa, Gouki and Tajima, Satohiro and Komatsu, Hidehiko , title =. Cerebral Cortex , year =

    work page

  34. [34]

    and Arun, Sripati P

    Zhivago, Kalathupiriyan A. and Arun, Sripati P. , title =. Journal of Neurophysiology , year =

    work page

  35. [35]

    and Tarr, Michael J

    Henderson, Margaret M. and Tarr, Michael J. and Wehbe, Leila , title =. Journal of Neuroscience , year =

    work page

  36. [36]

    and Gardner, Justin L

    Jagadeesh, Akshay V. and Gardner, Justin L. , title =. Nature Neuroscience , year =

    work page

  37. [37]

    and Kietzmann, Tim C

    Storrs, Katherine R. and Kietzmann, Tim C. and Walther, Alexander and Mehrer, Johannes and Kriegeskorte, Nikolaus , title =. Journal of Cognitive Neuroscience , year =

    work page

  38. [38]

    and Geirhos, Robert , title =

    Wichmann, Felix A. and Geirhos, Robert , title =. Annual Review of Vision Science , year =

    work page

  39. [39]

    Model Metamers Reveal Divergent Invariances between Biological and Artificial Neural Networks , journal =

    Feather, Jenelle and Leclerc, Guillaume and M. Model Metamers Reveal Divergent Invariances between Biological and Artificial Neural Networks , journal =. 2023 , volume =

    work page 2023

  40. [40]

    and Brendel, Wieland , title =

    Geirhos, Robert and Narayanappa, Kantharaju and Mitzkus, Benjamin and Thieringer, Tizian and Bethge, Matthias and Wichmann, Felix A. and Brendel, Wieland , title =. Advances in Neural Information Processing Systems , year =

    work page

  41. [41]

    and Rajalingham, Rishi and Issa, Elias B

    Schrimpf, Martin and Kubilius, Jonas and Hong, Ha and Majaj, Najib J. and Rajalingham, Rishi and Issa, Elias B. and Kar, Kohitij and Bashivan, Pouya and Prescott-Roy, Jonathan and Schmidt, Kailyn and Yamins, Daniel L. K. and DiCarlo, James J. , title =. bioRxiv , year =

    work page

  42. [42]

    Yamins, Daniel L. K. and Hong, Ha and Cadieu, Charles F. and Solomon, Ethan A. and Seibert, Darren and DiCarlo, James J. , title =. Proceedings of the National Academy of Sciences , year =

    work page

  43. [43]

    Proceedings of the 38th International Conference on Machine Learning , year =

    Radford, Alec and Kim, Jong Wook and Hallacy, Chris and Ramesh, Aditya and Goh, Gabriel and Agarwal, Sandhini and Sastry, Girish and Askell, Amanda and Mishkin, Pamela and Clark, Jack and Krueger, Gretchen and Sutskever, Ilya , title =. Proceedings of the 38th International Conference on Machine Learning , year =

    work page

  44. [44]

    From BoW to CNN: Two Decades of Texture Representation for Texture Classification , journal =

    Liu, Li and Chen, Jie and Fieguth, Paul and Zhao, Guoying and Chellappa, Rama and Pietik. From BoW to CNN: Two Decades of Texture Representation for Texture Classification , journal =. 2019 , volume =

    work page 2019

  45. [45]

    Wallis, Thomas S. A. and Funke, Christina M. and Ecker, Alexander S. and Gatys, Leon A. and Wichmann, Felix A. and Bethge, Matthias , title =. Journal of Vision , year =

    work page

  46. [46]

    , title =

    Bashivan, Pouya and Kar, Kohitij and DiCarlo, James J. , title =. Science , year =

    work page

  47. [47]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    work page

  48. [48]

    and Fei-Fei, Li , title =

    Russakovsky, Olga and Deng, Jia and Su, Hao and Krause, Jonathan and Satheesh, Sanjeev and Ma, Sean and Huang, Zhiheng and Karpathy, Andrej and Khosla, Aditya and Bernstein, Michael and Berg, Alexander C. and Fei-Fei, Li , title =. International Journal of Computer Vision , year =

    work page

  49. [49]

    International Conference on Learning Representations (ICLR) , year =

    Simonyan, Karen and Zisserman, Andrew , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  50. [50]

    DINOv2: Learning Robust Visual Features without Supervision , journal =

    Oquab, Maxime and Darcet, Timoth. DINOv2: Learning Robust Visual Features without Supervision , journal =. 2024 , note =

    work page 2024

  51. [51]

    Proceedings of the 37th International Conference on Machine Learning , year =

    Chen, Mark and Radford, Alec and Child, Rewon and Wu, Jeff and Jun, Heewoo and Dhariwal, Prafulla and Luan, David and Sutskever, Ilya , title =. Proceedings of the 37th International Conference on Machine Learning , year =

    work page

  52. [52]

    PNAS Nexus , year =

    Glielmo, Aldo and Zeni, Claudio and Cheng, Bingqing and Csanyi, Gabor and Laio, Alessandro , title =. PNAS Nexus , year =

    work page

  53. [53]

    Perceptual misalignment of texture representations in convolutional neural networks

    de Paolis, Ludovica and Anselmi, Fabio and Ansuini, Alessio and Piasini, Eugenio , title =. arXiv preprint arXiv:2604.01341 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  54. [54]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    Prefixing Attention Sinks can Mitigate Activation Outliers for Large Language Model Quantization , author =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    work page 2024

  55. [55]

    International Conference on Learning Representations (ICLR) , year =

    From Attention to Activation: Unraveling the Enigmas of Large Language Models , author =. International Conference on Learning Representations (ICLR) , year =

    work page

  56. [56]

    International Conference on Learning Representations (ICLR) , year =

    See What You Are Told: Visual Attention Sink in Large Multimodal Models , author =. International Conference on Learning Representations (ICLR) , year =

    work page

  57. [57]

    Ravishankar and Lohse, Gerald L

    Bhushan, Nalini and Rao, A. Ravishankar and Lohse, Gerald L. , title = ". Cognitive Science, 21. 1997

    work page 1997

  58. [58]

    Self-Guided Contrastive Learning for BERT Sentence Representations

    Kim, Taeuk and Yoo, Kang Min and Lee, Sang-goo. Self-Guided Contrastive Learning for BERT Sentence Representations. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021. doi:10.18653/v1/2021.acl-long.197

    work page doi:10.18653/v1/2021.acl-long.197 2021

  59. [59]

    Universal Text Representation from BERT: An Empirical Study

    Ma, Xiaofei and Wang, Zhiguo and Ng, Patrick and Nallapati, Ramesh and Xiang, Bing , title =. doi:10.48550/arXiv.1910.07973 , year =. 1910.07973 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1910.07973 1910

  60. [60]

    The geometry of hidden representations of large transformer models

    Valeriani, Lucrezia and Doimo, Diego and Cuturello, Francesca and Laio, Alessandro and Ansuini, Alessio and Cazzaniga, Alberto , title =. Advances in Neural Information Processing Systems , year =. 2302.00294 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv

  61. [61]

    doi:10.48550/arXiv.2601.21621 , year =

    Mahaut, Mat. doi:10.48550/arXiv.2601.21621 , year =. 2601.21621 , archivePrefix =

    work page doi:10.48550/arxiv.2601.21621

  62. [62]

    doi:doi:10.5281/zenodo.5561807 , url =

    Piasini, Eugenio , title =. doi:doi:10.5281/zenodo.5561807 , url =

    work page doi:10.5281/zenodo.5561807

  63. [63]

    , title =

    Portilla, Javier and Simoncelli, Eero P. , title =. 2000 , note =

    work page 2000

  64. [64]

    2026 , eprint =

    Acevedo, Santiago and Laio, Alessandro and Baroni, Marco , title =. 2026 , eprint =

    work page 2026

  65. [65]

    2022 , issn =

    DADApy: Distance-based analysis of data-manifolds in Python , journal =. 2022 , issn =. doi:https://doi.org/10.1016/j.patter.2022.100589 , url =

    work page doi:10.1016/j.patter.2022.100589 2022

  66. [66]

    ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness

    Geirhos, Robert and Rubisch, Patricia and Michaelis, Claudio and Bethge, Matthias and Wichmann, Felix A. and Brendel, Wieland , date =. doi:10.48550/ARXIV.1811.12231 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1811.12231

  67. [67]

    The Origins and Prevalence of Texture Bias in Convolutional Neural Networks , booktitle =

    Hermann, Katherine and Chen, Ting and Kornblith, Simon , editor =. The Origins and Prevalence of Texture Bias in Convolutional Neural Networks , booktitle =

    work page

  68. [68]

    Shape or Texture: Understanding Discriminative Features in CNNs

    Islam, Md Amirul and Kowal, Matthew and Esser, Patrick and Jia, Sen and Ommer, Bjorn and Derpanis, Konstantinos G. and Bruce, Neil , date =. Shape or. doi:10.48550/ARXIV.2101.11604 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2101.11604

  69. [69]

    Kubilius, Jonas and Schrimpf, Martin and Nayebi, Aran and Bear, Daniel and Yamins, Daniel L. K. and DiCarlo, James J. , date =. doi:10.1101/408385 , url =

    work page doi:10.1101/408385

  70. [70]

    Proceedings of the 41st International Conference on Machine Learning , pages =

    Position: The Platonic Representation Hypothesis , author =. Proceedings of the 41st International Conference on Machine Learning , pages =. 2024 , editor =

    work page 2024

  71. [71]

    Nature Neuroscience , year =

    Riesenhuber, Maximilian and Poggio, Tomaso , title =. Nature Neuroscience , year =. doi:10.1038/81479 , pmid =

    work page doi:10.1038/81479

  72. [72]

    Biological Cybernetics , year =

    Fukushima, Kunihiko , title =. Biological Cybernetics , year =. doi:10.1007/BF00344251 , pmid =

    work page doi:10.1007/bf00344251

  73. [73]

    and Wiesel, Torsten N

    Hubel, David H. and Wiesel, Torsten N. , title =. The Journal of Physiology , year =. doi:10.1113/jphysiol.1962.sp006837 , pmid =

    work page doi:10.1113/jphysiol.1962.sp006837 1962

  74. [74]

    , title =

    Lindsay, Grace W. , title =. Journal of Cognitive Neuroscience , year =. doi:10.1162/jocn_a_01544 , pmid =

    work page doi:10.1162/jocn_a_01544

  75. [75]

    , title =

    Krizhevsky, Alex and Sutskever, Ilya and Hinton, Geoffrey E. , title =. Advances in Neural Information Processing Systems , editor =

    work page

  76. [76]

    Annual Review of Vision Science , year =

    Kriegeskorte, Nikolaus , title =. Annual Review of Vision Science , year =

    work page

  77. [77]

    , title =

    Lindsay, Grace W. , title =. Journal of Cognitive Neuroscience , year =

    work page

  78. [78]

    , title =

    Kriegeskorte, Nikolaus and Mur, Marieke and Bandettini, Peter A. , title =. Frontiers in Systems Neuroscience , year =

    work page

  79. [79]

    International Conference on Learning Representations , year =

    Yu, Jiahui and Li, Xin and Koh, Jing Yu and Zhang, Han and Pang, Ruoming and Qin, James and Ku, Alexander and Xu, Yuanzhong and Baldridge, Jason and Wu, Yonghui , title =. International Conference on Learning Representations , year =

    work page

  80. [80]

    2026 , eprint=

    A quantitative analysis of semantic information in deep representations of text and images , author=. 2026 , eprint=

    work page 2026

Showing first 80 references.