Pith. sign in

REVIEW 4 major objections 6 minor 37 references

Foundations and Models in Modern Computer Vision: Key Building Blocks in Landmark Architectures

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This review argues that six landmark papers — ResNet, ViT, GAN, Latent Diffusion, DINO, and MAE — form one continuous story in which each design pattern answers a limitation of the previous, converging on self-supervised transformer-based…

desk verdict Readable survey of six classic vision papers; the summaries hold up, but the arbitrary selection and 'cohesive narrative' claim need rework. read the letter →

arxiv 2507.23357 v2 pith:YYPIP7RC submitted 2025-07-31 cs.CV

Radu-Andrei Bourceanu , Neil De La Fuente , Jan Grimm , Andrei Jardan , Andriy Manucharyan , Cornelius Weiss , Daniel Cremers , Roman Pflugfelder This is my paper · ORCID
classification cs.CV
keywords computervisionResNetTransformergenerativeadversarialnetworkslatentdiffusionmodelsself-supervisedlearningDINOmaskedautoencoders
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

This report is a technical review of six papers it treats as the critical milestones of modern computer vision: ResNet, the Vision Transformer, GANs, Latent Diffusion Models, DINO, and Masked Autoencoders. Its central claim is that these works form a single, continuous narrative, moving from convolutional backbones to attention-based transformers, then to generative modeling, and finally to self-supervised learning that removes the need for human labels. Each new method is presented as an answer to a limitation of the previous one, for example ViT's data hunger being addressed by DINO and MAE. A sympathetic reader would take the report as a distillation of the design patterns that now underlie both image recognition and image generation.

What carries the argument

The carrying object is the three-phase narrative itself, organised around six named mechanisms: the residual block (adding a layer's input to its output so gradients flow through deep networks), the vision transformer's image patch tokenization with global self-attention, the generator–discriminator minimax game of GANs, the latent diffusion denoising U-Net operating in a pre-trained autoencoder's compressed space, DINO's student–teacher self-distillation with centering and sharpening, and MAE's asymmetric encoder–decoder with roughly 75 percent random masking. Each mechanism is described in terms of the problem it solves and the empirical result that supports it.

What would settle it

A concrete check would be a citation or impact analysis of the field's history to see whether ResNet, ViT, GAN, Latent Diffusion, DINO, and MAE are genuinely the most influential works of their phases, or a re-run of one reported result, such as MAE's 87.8 percent ImageNet top-1 accuracy with a ViT-H, to see whether the numbers the narrative depends on actually hold.

Watch

Extended reading notes

Core claim

The paper's central discovery, stated in its conclusions, is that the trajectory of these six papers reveals a clear and cohesive narrative in computer vision research. It claims that the residual connection, patch-based self-attention, adversarial training, latent-space denoising, self-distillation, and masked reconstruction are successive design responses to the same underlying problems of depth, data, and label dependence. The report argues that these paradigms are not independent but converging, and that the field is moving toward massive, self-supervised, transformer-based foundation models serving both high-level recognition and high-fidelity controllable generation.

Load-bearing premise

The whole narrative rests on the selection assumption that these six papers are the critical milestones of computer vision; if other influential works had been included, the story could look different.

Editorial extensions

If this is right

  • Residual connections remain the default scaffolding for deep vision networks, and ViT blocks inherit the same skip-pathway idea despite abandoning convolutions.
  • Transformer-based vision models are data-hungry; self-supervised pretexts such as DINO's distillation and MAE's masking provide the inductive bias that CNN architectures had built in.
  • Latent diffusion is presented as the current efficient route to high-fidelity image generation, with remaining weaknesses in high-frequency details, memorization, and deepfake risk.
  • Self-supervised pre-training with DINO and MAE is claimed to close the label bottleneck, achieving strong classification and transfer results with no human annotations.
  • The report predicts that these strands converge into unified, self-supervised, transformer-based foundation models for both recognition and generation.

Reading between the lines

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

  • The narrative is retrospective: choosing a different set of influential works, for instance U-Net for segmentation or contrastive methods like SimCLR for self-supervision, would likely produce a different story, so the 'cohesive' arc is partly a product of the selection.
  • The reported numeric results are inherited from the original papers, not independently verified; readers should treat headline accuracies as claims from the primary sources.
  • A testable extension would quantify the claimed trajectory with citation or benchmark data across the field's history; if contrastive self-supervised methods dominate the third phase, the distillation-versus-reconstruction framing would need revision.
  • The convergence claim makes a concrete prediction: future multimodal foundation models will combine residual-style pathways, attention, latent denoising, and self-supervised pretexts in one backbone, which can be checked against upcoming architectures.
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

4 major / 6 minor

Summary. This manuscript is a survey of six influential computer vision papers: ResNet, the Vision Transformer (ViT), GANs, Latent Diffusion Models (LDMs), DINO, and Masked Autoencoders (MAE). For each paper, it summarizes the motivation, method, key experiments, and discussion, and it groups the six works into three phases: foundational backbones, generative modeling, and self-supervised representation learning. The survey argues in the conclusions that these six papers reveal a clear and cohesive narrative of computer vision's evolution.

Significance. The paper's value is pedagogical: it condenses six well-known papers and reproduces their core formulations, including the GAN min-max objective, DINO's EMA/centering/sharpening equations, and the MAE masking pipeline. The individual summaries are mostly faithful to the original sources, which is a genuine strength for a survey. However, the manuscript makes no new technical contribution, does not run experiments or verify reported numbers, and its central narrative claim depends on a selection of papers that is never justified. The significance is therefore moderate and contingent on the selection being made explicit.

major comments (4)
  1. [§1, §3.2.1–3.2.2, §5] The manuscript's central narrative claim in Section 5—that 'the trajectory of progress shown by these six papers reveals a clear and cohesive narrative'—rests on the unsupported premise in Section 1 that these are 'critical milestones.' No inclusion criteria are provided, and internal evidence weakens the closed-set premise: Section 3.2.1 and Section 3.2.2 describe the LDM denoising network as a U-Net and rely on a pre-trained VAE-style autoencoder for perceptual compression, yet neither component is analyzed in Section 2 or counted among the six papers. The authors should make the selection criteria explicit, or soften the 'cohesive narrative' claim to acknowledge that the chosen subset is one possible genealogy rather than the unique trajectory.
  2. [§3.1.2] The GAN training procedure is described in reverse: the text says training alternates 'k steps for updating the generator and one step for the discriminator.' In the original GAN paper (Algorithm 1), k steps update the discriminator for each generator step, which is consistent with the preceding sentence about keeping D close to optimal. This inversion is material to the exposition of adversarial training dynamics and should be corrected.
  3. [Abstract, §3.2.3, §5] The claim that LDMs are 'the current state-of-the-art for image generation' is unsupported as written. The numerical evidence cited in Section 3.2.3 is contemporaneous to the 2022 LDM paper (e.g., parity with Make-A-Scene and GLIDE on MS-COCO), which does not license a 'current' state-of-the-art statement in a 2025 survey. Please replace 'current state-of-the-art' with 'state-of-the-art at the time of publication' or provide current comparative benchmark evidence.
  4. [§4.2.1, §4.2.4] The MAE sections state that MAE 'leads to SOTA results' and 'outperforms contrastive pre-training methods on classification tasks' without noting an important caveat from the original MAE paper: linear probing of MAE features is substantially weaker than that of contrastive methods. Since the survey elsewhere emphasizes k-NN and linear evaluation in the DINO section, this asymmetry should be acknowledged to avoid overstating the comparison.
minor comments (6)
  1. [§2.1.1] Reference [22] is a Microsoft blog rather than the ResNet paper; the claim about ResNet's introduction should cite He et al. [14] directly.
  2. [§2.2.1, §4.1.1] GPT is cited as [24], which is an entry for ChatGPT rather than the GPT paper; please replace it with the appropriate Radford et al. or Brown et al. reference.
  3. [§2.2 heading] The section heading contains a typo: 'Vision T ransformer' should be 'Vision Transformer.'
  4. [§3.1.5] There is a missing space in 'LSUN bedrooms256×256 px'; also, the citation for ProGAN should be formatted consistently with the other references.
  5. [References] References [9] and [22] are non-scholarly web sources; for a survey, consider replacing them with peer-reviewed or preprint sources.
  6. [§4.1.4] The phrase 'self-attention maps look remarkably like segmentation maps' would be more precise as 'segmentation-like maps,' since DINO-trained ViTs are not trained for segmentation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the report is a survey restating external results, with an interpretive selection claim rather than a derivation.

full rationale

This manuscript is a literature survey that summarizes six named external papers (ResNet, ViT, GAN, LDM, DINO, MAE). It performs no new derivation, fits no parameters, and does not define one of its objects in terms of another. The central sentence in the Conclusions, namely that 'the trajectory of progress shown by these six papers reveals a clear and cohesive narrative,' is an editorial interpretation of the selected works, not a result derived from those works. The Introduction's assertion that the six papers 'represent critical milestones' is a selection decision without a formal criterion, so it is unsupported as a claim of representativeness, but it is not circular: the survey does not use the conclusion of cohesiveness as evidence for the selection, nor does it use the selection as evidence for the conclusion in a way that closes a logical loop. The discussion of LDMs mentions that the denoising network is a U-Net, yet U-Net is not among the six surveyed papers. This is a completeness or scope limitation, not a circularity: the survey is not claiming to derive U-Net's role from the selected set, and the omission does not make any stated result equivalent to its input. There are no self-citations to the present authors' prior work, no imported uniqueness theorem, and no fitted quantity renamed as a prediction. The reported experimental numbers are taken from the original papers without re-running them, which is a verification limitation but again not circular reasoning. Under the stated criteria, no circular step can be quoted and exhibited, so the appropriate finding is no significant circularity with score 0.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

No free parameters or invented entities are introduced. The report relies on two domain assumptions: representativeness of the six selected papers and accuracy of the cited experimental results.

assumptions (2)
  • domain assumption The six selected papers are representative of the three major phases in computer vision evolution.
    The Introduction states these six papers represent critical milestones, but no justification is given for excluding other influential works. The Conclusions build a cohesive narrative on this selection.
  • domain assumption The experimental results and numbers reported from the cited papers are accurate.
    The survey does not re-run any experiments. It trusts the results as reported in the original papers, which are external and not verified here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Foundations and Models in Modern Computer Vision: Key Building Blocks in Landmark Architectures." pith.science (2026). https://pith.science/paper/YYPIP7RC

@misc{pith2026250723357,
  author       = {Pith},
  title        = {Pith review of: Foundations and Models in Modern Computer Vision: Key Building Blocks in Landmark Architectures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YYPIP7RC}},
  note         = {Machine review of arXiv:2507.23357}
}
read the original abstract

This report analyzes the evolution of key design patterns in computer vision by examining six influential papers. The analysis begins with foundational architectures for image recognition. We review ResNet, which introduced residual connections to overcome the vanishing gradient problem and enable effective training of significantly deeper convolutional networks. Subsequently, we examine the Vision Transformer (ViT), which established a new paradigm by applying the Transformer architecture to sequences of image patches, demonstrating the efficacy of attention-based models for large-scale image recognition. Building on these visual representation backbones, we investigate generative models. Generative Adversarial Networks (GANs) are analyzed for their novel adversarial training process, which challenges a generator against a discriminator to learn complex data distributions. Then, Latent Diffusion Models (LDMs) are covered, which improve upon prior generative methods by performing a sequential denoising process in a perceptually compressed latent space. LDMs achieve high-fidelity synthesis with greater computational efficiency, representing the current state-of-the-art for image generation. Finally, we explore self-supervised learning techniques that reduce dependency on labeled data. DINO is a self-distillation framework in which a student network learns to match the output of a momentum-updated teacher, yielding features with strong k-NN classification performance. We conclude with Masked Autoencoders (MAE), which utilize an asymmetric encoder-decoder design to reconstruct heavily masked inputs, providing a highly scalable and effective method for pre-training large-scale vision models.

Figures

Figures reproduced from arXiv: 2507.23357 by the authors.

Figure 1
Figure 1. Residual learning: a building block. In courtesy of He et al. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. One variant of a residual block. Over time, several variants of the residual block have been proposed and adopted in practice. For example, placing Batch Normalization after the addition, instead of after each convolution, has shown improved stability in some contexts. Addi￾tionally, some versions omit the final ReLU activation after the addition, allowing the identity mapping to pass through more directly. For thei… view at source ↗
Figure 3
Figure 3. Training on ImageNet. The authors of Deep Residual Learning for Image Recognition performed experiments to evaluate the effectiveness of resid￾ual learning on large-scale image classification tasks. A key finding was the impact of residual connections on network optimization, especially when increasing depth. In one experiment, a plain convolutional network with 34 layers was directly compared to a shallower 18-laye… view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: Error rates (%) on the ImageNet validation ( [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Vision Transformer Architecture [10]. In courtesy of the authors. To apply this idea to images, the image must first be sequentialized so that it can be accepted by the transformer encoder. This is done by dividing the input image into small, fixed-size patches, such a…
Figure 6
Figure 6. Figure 6: Comparison of Vision Transformer and CNNs on multiple image classification datasets. [ [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Comparison of Vision Transformer and ResNet on different training dataset sizes. [ [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Overview of the vanilla GAN framework. To illustrate training dynamics, the authors visualize how the generator distribution pg, the discriminator output D(x), as well as the true data distribution pdata evolve over the course of training. After random initialization, …
Figure 9
Figure 9. Figure 9: Visualization of distributions during GAN training. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Diffusion Model architecture. 3.2.2 Approach and Methodology Latent Diffusion Models promise to solve this problem by moving the training and image generation from the high-dimensional pixel space into a lower-dimensional latent space. This is achieved through two mai…
Figure 11
Figure 11. Figure 11: Latent Diffusion Model architecture. In courtesy of [ [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: DINO illustration local and global views (large crops) is extracted. The set x1 is passed through the student model, followed by a softmax operation, yielding the student’s probability dis￾tribution Ps. The set x2 is passed through the teacher model, followed by a cen…
Figure 13
Figure 13. Figure 13: Comparison of accuracy achieved with differently-constructed teachers. [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: Accuracies of student and teacher throughout training. [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]
Figure 15
Figure 15. Figure 15: Left: evolution of entropy of teacher’s output; Right: evolution [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]
Figure 16
Figure 16. Figure 16: Self-attention from a Vision Transformer trained under DINO. In courtesy of the authors in [ [PITH_FULL_IMAGE:figures/full_fig_p020_16.png]
Figure 17
Figure 17. Figure 17: MAE architecture. In courtesy of [12]. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_17.png]
Figure 18
Figure 18. Figure 18: Example results on ImageNet validation images. For each triplet, the masked image (left), the MAE [PITH_FULL_IMAGE:figures/full_fig_p023_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 25 canonical work pages

  1. [22]

    Microsoft researchers win ImageNet computer vision challenge

    Allison Linn. Microsoft researchers win ImageNet computer vision challenge. https://blogs.microsoft. com/ai/microsoft-researchers-win-imagenet-computer-vision-challenge/ . Accessed: 2025-07-16

    work page 2025

  2. [14]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  3. [1]

    Bert: A sentiment analysis odyssey

    Shivaji Alaparthi and Manit Mishra. Bert: A sentiment analysis odyssey. Journal of Marketing Analytics, 9(2):118–126, 2021

    work page 2021

  4. [2]

    Beit: Bert pre-training of image transformers

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021

    arXiv 2021

  5. [3]

    Large scale gan training for high fidelity natural image synthesis, 2019

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis, 2019

    work page 2019

  6. [4]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J ´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650–9660, 2021

    work page 2021

  7. [5]

    A simple framework for contrastive learning of visual representations, 2020

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations, 2020

    2020

  8. [6]

    An empirical study of training self-supervised vision transformers

    Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640–9649, 2021

    work page 2021

Show all 37 references
  1. [7]

    A comparative study between vision transformers and cnns in digital pathology

    Luca Deininger, Bernhard Stimpel, Anil Yuce, Samaneh Abbasi-Sureshjani, Simon Sch ¨onenberger, Paolo Ocampo, Konstanty Korski, and Fabien Gaire. A comparative study between vision transformers and cnns in digital pathology. arXiv preprint arXiv:2206.00389, 2022

    2022 arXiv

  2. [8]

    BERT: Pre-training of deep bidirectional trans- formers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional trans- formers for language understanding. In Jill Burstein, Christy Doran, and Thamar Solorio, editors, Proceedings of the 2019 Conference of the North American Chapter ...

    2019

  3. [9]

    Gan vs transformer: A generative ai comparison, June 2025

    DhiWise. Gan vs transformer: A generative ai comparison, June 2025. Accessed: 2025-07-05

    2025

  4. [10]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arX...

    2010 arXiv

  5. [11]

    Generative adversarial nets

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 25

    2014

  6. [12]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000– 16009, 2022

    2022

  7. [13]

    Momentum contrast for unsupervised visual representation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020

    2020

  8. [15]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018

    2018

  9. [16]

    Denoising diffusion probabilistic models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020

    2020

  10. [17]

    Long short-term memory

    Sepp Hochreiter and J ¨urgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735–1780, November 1997

    1997

  11. [18]

    Progressive growing of gans for improved quality, stability, and variation, 2018

    Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation, 2018

    2018

  12. [19]

    Big transfer (bit): General visual representation learning

    Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In European conference on computer vision, pages 491–507. Springer, 2020

    2020

  13. [20]

    Imagenet classification with deep convolutional neural net- works

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net- works. Advances in neural information processing systems, 25, 2012

    2012

  14. [21]

    Efficient self-supervised vision transformers for representation learning, 2022

    Chunyuan Li, Jianwei Yang, Pengchuan Zhang, Mei Gao, Bin Xiao, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Efficient self-supervised vision transformers for representation learning, 2022

    2022

  15. [23]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021

    2021

  16. [24]

    ChatGPT, 2025

    OpenAI. ChatGPT, 2025. Large language model, accessed 2025-07-20

    2025

  17. [25]

    Dinov2: Learning robust visual features without supervision, 2024

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael ...

    2024

  18. [26]

    High-resolution image syn- thesis with latent diffusion models, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image syn- thesis with latent diffusion models, 2022

    2022

  19. [27]

    Emre Celebi, and Jie Yang

    Pourya Shamsolmoali, Masoumeh Zareapoor, Eric Granger, Huiyu Zhou, Ruili Wang, M. Emre Celebi, and Jie Yang. Image synthesis with adversarial networks: a comprehensive survey and case studies, 2020

    2020

  20. [28]

    Very deep convolutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. Septem- ber 2014

    2014

  21. [29]

    Highway networks

    Rupesh Kumar Srivastava, Klaus Greff, and J ¨urgen Schmidhuber. Highway networks. CoRR, abs/1505.00387, 2015. 26

    2015 arXiv

  22. [30]

    Going deeper with convolutions, 2014

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions, 2014

    2014

  23. [31]

    Are convolutional neural networks or transformers more like human vision? arXiv preprint arXiv:2105.07197, 2021

    Shikhar Tuli, Ishita Dasgupta, Erin Grant, and Thomas L Griffiths. Are convolutional neural networks or transformers more like human vision? arXiv preprint arXiv:2105.07197, 2021

    2021 arXiv

  24. [32]

    Attention is all you need, 06 2017

    Ashish Vaswani, Google Brain, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need, 06 2017

    2017

  25. [33]

    Non-local neural networks

    Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794–7803, 2018

    2018

  26. [34]

    Scaling vision transformers

    Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104–12113, 2022

    2022

  27. [35]

    Progressive augmentation of gans, 2019

    Dan Zhang and Anna Khoreva. Progressive augmentation of gans, 2019

    2019

  28. [36]

    Energy-based generative adversarial network, 2017

    Junbo Zhao, Michael Mathieu, and Yann LeCun. Energy-based generative adversarial network, 2017

    2017

  29. [37]

    ibot: Image bert pre-training with online tokenizer, 2022

    Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer, 2022. 27

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.