# Image classification

> Source: https://aiwiki.ai/wiki/image_classification
> Updated: 2026-07-24
> Fact-checked: 2026-07-24
> Categories: AI Benchmarks, Computer Vision, Deep Learning, Machine Learning
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Image classification." aiwiki.ai, 24 Jul 2026. https://aiwiki.ai/wiki/image_classification
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Image classification is the task of assigning a whole image to one category drawn from a fixed set of labels. The input is a single picture, the output is one label or a ranked list of candidate labels, and nothing is said about where in the frame the object sits. That distinguishes it from [object detection](https://aiwiki.ai/wiki/object_detection), which localizes instances with bounding boxes, and from [image segmentation](https://aiwiki.ai/wiki/image_segmentation), which assigns a label to every pixel.

The task became the proving ground for [deep learning](https://aiwiki.ai/wiki/deep_learning) in [computer vision](https://aiwiki.ai/wiki/computer_vision). Between 2010 and 2017 the annual ImageNet Large Scale Visual Recognition Challenge (ILSVRC) cut the top-5 error rate on a 1,000-category classification problem from 28.2% to 2.25% [1][24]. The 2012 winner, a [convolutional neural network](https://aiwiki.ai/wiki/convolutional_neural_network) now known as [AlexNet](https://aiwiki.ai/wiki/alexnet), is usually treated as the point at which the field abandoned hand-engineered features for learned ones: the challenge organizers noted that after 2012 the vast majority of entries switched to deep convolutional networks [1][3].

Classification also remains the standard way to compare visual backbones. A network trained to label [ImageNet](https://aiwiki.ai/wiki/imagenet) photographs is rarely deployed as a photo labeler; it is far more often stripped of its final layer and reused as a [feature extractor](https://aiwiki.ai/wiki/feature_extraction) for detection, segmentation, retrieval, or medical diagnosis [5][6]. Accuracy on ImageNet has therefore functioned less as a product metric than as a proxy for representation quality.

## How a classifier is built

A modern classifier is a function that maps pixels to a probability distribution over labels, trained end to end by [gradient descent](https://aiwiki.ai/wiki/gradient_descent). The standard pipeline has four stages.

Preprocessing resizes and crops each image to a fixed resolution and normalizes pixel values. AlexNet rescaled each image so that its shorter side was 256 pixels, cropped out the central 256x256 patch, and subtracted the training-set mean from each pixel, doing no other preprocessing [3]. Most subsequent work has kept this shape of pipeline while raising the resolution.

[Data augmentation](https://aiwiki.ai/wiki/data_augmentation) expands the effective training set with label-preserving transformations such as random crops, horizontal flips, and color jitter. LeCun and colleagues already used this trick in 1998, generating 540,000 distorted variants of the 60,000 original [MNIST](https://aiwiki.ai/wiki/mnist) digits [7]. The technique matters enough that later work raised a [ResNet](https://aiwiki.ai/wiki/resnet)-50 from 75.3% to 79.29% top-1 accuracy through training-procedure refinements such as augmentation and optimization changes [8], and a vanilla ResNet-50 to 80.4% at 224x224 resolution with a more demanding recipe and no extra data or distillation [9].

The network itself produces a vector of class scores, converted to probabilities by a [softmax](https://aiwiki.ai/wiki/softmax) and scored against the ground-truth label with a cross-entropy [loss function](https://aiwiki.ai/wiki/loss_function) [3]. [Dropout](https://aiwiki.ai/wiki/dropout) in the fully connected layers and, from 2015, [batch normalization](https://aiwiki.ai/wiki/batch_normalization) between layers were the two regularization and conditioning tricks that made deeper stacks trainable [3][10].

At inference the model returns the arg-max label, or the top few labels ranked by probability. Practitioners commonly evaluate several crops or scales of the same image and average the predictions: AlexNet already averaged its softmax outputs over ten patches, the four corners and the center of the image plus their horizontal reflections, which cut its top-5 error by more than a point [3].

## Measuring accuracy: top-1 and top-5

Two metrics dominate. Top-1 accuracy counts a prediction as correct only when the highest-scoring label matches the ground truth. Top-5 accuracy counts it as correct if the ground-truth label appears anywhere in the model's five highest-scoring guesses.

Top-5 exists because of an annotation problem rather than a modeling one. ImageNet gives each image a single class label, but real photographs contain several objects: an image labeled "strawberry" may also contain an apple, and a classifier naming the apple is not obviously wrong. The ILSVRC organizers therefore allowed an algorithm to return five labels and counted it correct if any of them matched [1]. Papers from the 2012-2017 era usually report top-5 error, while more recent work reports top-1 accuracy. Top-5 error had fallen to 2.25% by the final ILSVRC, with the runner-up at 2.48% [24].

Human performance on the same protocol has been measured directly. One trained annotator working through 1,500 ILSVRC test images reached 5.1% top-5 error, against 6.8% for GoogLeNet on the same sample. A second annotator, who practiced on only 100 images, managed roughly 12.0%, which led the organizers to conclude that a significant amount of training time is necessary for a human to be competitive on the task [1].

## Benchmark datasets

| Dataset | Classes | Images | Resolution | Origin |
|---|---|---|---|---|
| MNIST | 10 | 60,000 train, 10,000 test | 28x28 grayscale | Built from NIST Special Databases 1 and 3, digits written by Census Bureau employees and high-school students [7] |
| [CIFAR-10](https://aiwiki.ai/wiki/cifar_10) | 10 | 50,000 train, 10,000 test | 32x32 color | Labeled subset of the 80 million tiny images dataset, by Alex Krizhevsky, Vinod Nair and [Geoffrey Hinton](https://aiwiki.ai/wiki/geoffrey_hinton) [11] |
| CIFAR-100 | 100 (20 superclasses) | 500 train and 100 test per class | 32x32 color | Same collection as CIFAR-10, with fine and coarse labels [11] |
| ImageNet (full) | 21,841 synsets | 14,197,122 | Variable | Web images organized on the [WordNet](https://aiwiki.ai/wiki/wordnet) noun hierarchy, verified through [Amazon Mechanical Turk](https://aiwiki.ai/wiki/amazon_mechanical_turk) [12][13] |
| ILSVRC-2012 subset | 1,000 | 1,281,167 train, 50,000 val, 100,000 test | Variable | The competition subset, still the default meaning of "ImageNet" in benchmark tables [1] |

ImageNet began as a 2009 CVPR paper by Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li and [Fei-Fei Li](https://aiwiki.ai/wiki/fei_fei_li) at Princeton University. At publication it covered 12 subtrees containing 5,247 synsets and 3.2 million images, with the stated ambition of populating most of WordNet's roughly 80,000 noun synsets at 500-1,000 images each [12]. Images were pulled from search engines and then verified by crowd workers, an effort that eventually involved more than 50,000 workers and over 160 million candidate images [13].

## Before deep learning

Handwritten digit recognition was the earlier standard task. LeCun, Bottou, Bengio and Haffner's 1998 paper in Proceedings of the IEEE introduced [LeNet](https://aiwiki.ai/wiki/lenet)-5, a seven-layer convolutional network with 60,000 trainable parameters, and reported a 0.95% test error rate on MNIST [7]. Dan Cireşan, Ueli Meier and Juergen Schmidhuber later reached what they described as the first near-human performance on MNIST using GPU-trained multi-column deep networks [14].

On natural images, though, the dominant approach through 2011 was a pipeline of hand-designed local descriptors fed to a linear classifier. The first ILSVRC in 2010 was won by the NEC-UIUC team with SIFT and local binary pattern features, two non-linear coding representations and a stochastic support vector machine, at 28.19% top-5 error [15]. The 2011 winner was XRCE, using high-dimensional Fisher vector image signatures compressed with product quantization and one-versus-all linear SVMs, at 25.77% [1][16]. Fisher vector methods stayed competitive in every challenge from 2010 to 2014 [1].

## AlexNet and the 2012 result

The 2012 challenge was won by SuperVision, the team of Alex Krizhevsky, [Ilya Sutskever](https://aiwiki.ai/wiki/ilya_sutskever) and Geoffrey Hinton at the University of Toronto. Their network had five convolutional and three fully connected layers, 60 million parameters and 650,000 neurons, and was trained for five to six days on two NVIDIA GTX 580 3GB [GPUs](https://aiwiki.ai/wiki/gpu) [3].

The official results show their best entry at 15.315% top-5 error using extra training data from the ImageNet Fall 2011 release, and 16.422% using only the supplied training data. The next team, ISI, scored 26.172% [2]. The gap of roughly eleven points over a field of hand-engineered systems is what made the result decisive rather than incremental.

The paper attributed its performance to a handful of specific choices: rectified linear units instead of saturating nonlinearities, a GPU implementation of 2D convolution, and dropout in the fully connected layers to control overfitting. It also noted that removing any single convolutional layer, each holding under 1% of the model's parameters, degraded results, which was taken as evidence that depth itself mattered [3].

## Architectures after AlexNet

| Year | Winning classification entry | Top-5 error | Approach |
|---|---|---|---|
| 2010 | NEC-UIUC | 28.19% | SIFT and LBP features, stochastic SVM [15] |
| 2011 | XRCE | 25.77% | Fisher vectors, linear SVMs [16] |
| 2012 | SuperVision (AlexNet) | 15.32% | 8-layer CNN on two GPUs [2][3] |
| 2013 | Clarifai | 11.20% | CNN redesigned using deconvolutional visualization [17][18] |
| 2014 | [GoogLeNet](https://aiwiki.ai/wiki/googlenet) | 6.66% | 22-layer [Inception](https://aiwiki.ai/wiki/inception) architecture [19][20] |
| 2015 | MSRA (ResNet) | 3.57% | 152-layer residual network [22] |
| 2016 | Trimps-Soushen | 2.99% | Large model ensemble [23] |
| 2017 | WMW (SENet) | 2.25% | Squeeze-and-excitation blocks [24][25] |

The 2012 and 2013 winning entries used training data beyond the supplied set; restricted to the provided data they scored 16.42% and 11.74% respectively [2][17]. The challenge ran annually from 2010 through 2017 [4].

Matthew Zeiler and Rob Fergus introduced a deconvolutional visualization technique that showed what intermediate feature layers respond to, then used it to find architectures that outperformed AlexNet [18]. The winning 2013 classification entry came from Clarifai, submitted by Zeiler, at 11.197% top-5 error with an additional model trained on 5,000 categories, or 11.743% using only the original training data [17].

Two 2014 entries defined the next generation. [Karen Simonyan](https://aiwiki.ai/wiki/karen_simonyan) and Andrew Zisserman's [VGG](https://aiwiki.ai/wiki/vgg) networks showed that stacking very small 3x3 convolution filters to a depth of 16-19 weight layers produced a significant improvement, and took first place in localization and second in classification [26]. [Christian Szegedy](https://aiwiki.ai/wiki/christian_szegedy) and colleagues' GoogLeNet won classification with a 22-layer network whose Inception modules increased depth and width while holding the computational budget constant [19].

[Kaiming He](https://aiwiki.ai/wiki/kaiming_he), Xiangyu Zhang, Shaoqing Ren and [Jian Sun](https://aiwiki.ai/wiki/jian_sun) then removed depth as a barrier. Their residual learning framework reformulated layers to learn a residual function with respect to the layer input, letting them train networks of 152 layers, described as eight times deeper than VGG while still having lower complexity. An ensemble reached 3.57% error on the ImageNet test set and won ILSVRC 2015 across classification, detection and localization [22]. Residual connections have since become near-universal.

Several threads ran in parallel. Batch normalization reached the accuracy of an existing image classification model in 14 times fewer training steps, and an ensemble of batch-normalized networks reached 4.9% top-5 validation error [10]. He and colleagues' parametric ReLU networks hit 4.94% top-5 test error and were, by their account, the first to pass the 5.1% human figure [21]. [DenseNet](https://aiwiki.ai/wiki/densenet) connected every layer to every subsequent layer to improve gradient flow and parameter efficiency [27]. [MobileNets](https://aiwiki.ai/wiki/mobilenet) traded accuracy for latency using depthwise separable convolutions, aimed at [on-device](https://aiwiki.ai/wiki/on_device_ai) deployment [28].

Mingxing Tan and [Quoc Le](https://aiwiki.ai/wiki/quoc_le)'s [EfficientNet](https://aiwiki.ai/wiki/efficientnet) took a different angle, arguing that depth, width and input resolution should be scaled together by a single compound coefficient rather than one at a time. EfficientNet-B7 reached 84.3% top-1 accuracy while being, by their measurement, 8.4 times smaller and 6.1 times faster at inference than the best existing convolutional network [29]. Self-training pushed the same family further: Noisy Student, which trains a larger student on pseudo-labels generated by a teacher over 300 million unlabeled images, reached 88.4% top-1 [30].

## Transfer learning

[Transfer learning](https://aiwiki.ai/wiki/transfer_learning), meaning the reuse of an ImageNet-trained network on a different dataset, is the normal way to build a classifier when labels are scarce. Jason Yosinski, Jeff Clune, [Yoshua Bengio](https://aiwiki.ai/wiki/yoshua_bengio) and Hod Lipson quantified why this works: early layers learn generic Gabor-like filters and color blobs that transfer well, transferability declines as the distance between source and target task grows, and initializing from transferred features gives a generalization boost that persists even after [fine-tuning](https://aiwiki.ai/wiki/fine_tuning) [5].

Simon Kornblith, Jonathon Shlens and Quoc Le tested the implicit assumption that better ImageNet models transfer better. Across 16 architectures and 12 datasets they found a strong correlation (r = 0.99 as fixed feature extractors, 0.96 fine-tuned), but with two caveats: regularization tricks that slightly improve ImageNet accuracy can badly damage penultimate-layer features for transfer, and on some fine-grained datasets ImageNet pretraining provided minimal benefit [6].

Pretraining on the larger ImageNet-21K rather than the 1,000-class subset generally helps further, though for years the larger set was awkward enough to prepare that most groups skipped it [31].

## Vision transformers

Alexey Dosovitskiy and colleagues at Google showed in October 2020 that convolutions were not required. The [Vision Transformer](https://aiwiki.ai/wiki/vision_transformer) splits an image into fixed-size patches, embeds each patch linearly, and feeds the sequence to a standard transformer encoder. Trained on small datasets it underperforms convolutional networks, because it lacks their built-in assumptions about locality and translation invariance; trained on ImageNet-21k or the in-house JFT-300M collection it overtakes them. Their best model reached 88.55% on ImageNet and 94.55% on CIFAR-100, and did so with 2,500 TPUv3-core-days against 9,900 for the Big Transfer baseline and 12,300 for Noisy Student [32].

The architectural argument did not end there. [Swin Transformer](https://aiwiki.ai/wiki/swin_transformer) reintroduced hierarchy and locality through shifted-window attention, reaching 87.3% top-1 on ImageNet-1K while working better as a general backbone [33]. Zhuang Liu and colleagues then modernized a plain ResNet step by step toward transformer design conventions and produced [ConvNeXt](https://aiwiki.ai/wiki/convnext), a pure convolutional network reaching 87.8% [34]. The comparison suggested that much of the transformer advantage came from training recipes and scale rather than attention itself.

## Language supervision and zero-shot classification

Every system above requires the label set to be fixed before training. [CLIP](https://aiwiki.ai/wiki/clip), described by [Alec Radford](https://aiwiki.ai/wiki/alec_radford) and colleagues at [OpenAI](https://aiwiki.ai/wiki/openai) in a paper posted in February 2021, broke that constraint by training an image encoder and a text encoder jointly on 400 million image-text pairs scraped from the web, using a contrastive objective that matches each image to its caption. Classification then becomes a retrieval problem: encode the candidate class names as text, encode the image, and pick the nearest. CLIP reached 76.2% top-1 and 95% top-5 zero-shot on ImageNet, matching the original ResNet-50 without using any of the 1,281,167 labeled training images [35].

Later work extended the recipe. CoCa combined contrastive and captioning losses and reported 86.3% zero-shot, 90.6% with a frozen encoder and a learned head, and 91.0% fine-tuned [36]. [SigLIP](https://aiwiki.ai/wiki/siglip) 2, released in February 2025 by a team at Google, added captioning-based pretraining, self-supervised losses such as self-distillation and masked prediction, and online data curation, and outperformed its predecessor at every model scale on core capabilities including zero-shot classification [37].

Purely self-supervised training caught up as well. [DINOv2](https://aiwiki.ai/wiki/dinov2) trained a 1B-parameter transformer on curated unlabeled images and distilled it into smaller models [38]. [DINOv3](https://aiwiki.ai/wiki/dinov3), released by Meta in August 2025, reached 88.4% ImageNet validation top-1 under a linear probe on frozen features, which its authors call the first time a self-supervised model has reached results comparable to weakly and fully supervised models on image classification [39].

## Robustness, bias, and the limits of the benchmark

High ImageNet accuracy has repeatedly turned out to be narrower than it looks.

Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt and Vaishaal Shankar rebuilt the ImageNet and CIFAR-10 test sets by following the original collection process, and found accuracy drops of 11-14% on ImageNet and 3-15% on CIFAR-10. The drops were not caused by test-set reuse, they concluded, but by models failing to generalize to slightly harder images from the same nominal distribution [40]. Lucas Beyer and coauthors collected more careful human annotations for the ImageNet validation set and found the gains of recent classifiers to be substantially smaller under the new labels, and the original labels to be no longer the best predictor of the independently collected ones [41].

Robustness benchmarks tell a similar story. ImageNet-C measures accuracy under common corruptions such as blur and noise, and its authors found essentially no relative improvement in corruption robustness from AlexNet to ResNet [42]. ImageNet-A, built by filtering for natural images that classifiers get wrong, reduced a DenseNet-121 to roughly 2% accuracy [43]. Robert Geirhos and colleagues showed that ImageNet-trained networks classify mainly by texture rather than shape, in contrast with human observers, and that retraining on stylized images shifts the bias and improves robustness [44]. Christian Szegedy and coauthors had already shown in 2013 that imperceptible, deliberately computed perturbations flip predictions, and that the same perturbation can cause a network trained on a different subset of the data to misclassify the same input [45].

The dataset itself has been the subject of ethical revision. Kaiyu Yang, Klint Qinami, Fei-Fei Li, Jia Deng and Olga Russakovsky examined the person subtree of ImageNet, identified 1,593 of its 2,832 synsets as offensive or sensitive, and began preparing a filtered version; downloads of the full dataset were disabled in January 2019 while the work proceeded, leaving only the 1,000-class challenge subset available [13][46]. That subset contains only three people categories: scuba diver, bridegroom and baseball player [13].

## Applications

Two medical results are commonly cited as evidence that the technique transfers to consequential decisions. Andre Esteva and colleagues trained a single convolutional network on 129,450 clinical images spanning 2,032 diseases and reported performance on par with 21 board-certified dermatologists on two binary skin-cancer tasks [47]. Varun Gulshan and colleagues trained a network on 128,175 retinal fundus photographs, each graded three to seven times by a panel of 54 US-licensed ophthalmologists and ophthalmology senior residents, and validated an algorithm for detecting referable diabetic retinopathy [48]. Both are classification problems in the strict sense: one image in, one label out.

The same formulation underlies content moderation, product and plant identification, industrial inspection, satellite and aerial land-cover mapping, and the perception stacks of [autonomous vehicles](https://aiwiki.ai/wiki/autonomous_vehicle), where classification heads sit downstream of detection. Reference implementations distributed with [PyTorch](https://aiwiki.ai/wiki/pytorch) make the accuracy and cost tradeoff explicit:

| Model | Top-1 | Top-5 | Parameters | GFLOPS |
|---|---|---|---|---|
| AlexNet | 56.52 | 79.07 | 61.1M | 0.71 |
| GoogLeNet | 69.78 | 89.53 | 6.6M | 1.50 |
| VGG-16 | 71.59 | 90.38 | 138.4M | 15.47 |
| MobileNet V3 Large | 74.04 | 91.34 | 5.5M | 0.22 |
| ResNet-50 | 76.13 | 92.86 | 25.6M | 4.09 |
| EfficientNet-B0 | 77.69 | 93.53 | 5.3M | 0.39 |
| ViT-B/16 | 81.07 | 95.32 | 86.6M | 17.56 |
| ConvNeXt Large | 84.41 | 96.98 | 197.8M | 34.36 |

Figures are torchvision's own reference weights evaluated on the ImageNet-1K validation set [49]. The MobileNet row shows the point: MobileNet V3 Large beats VGG-16 on top-1 accuracy with roughly one seventieth of the FLOPs and one twenty-fifth of the parameters.

## Recent developments

Headline ImageNet accuracy has moved slowly. A ViT-G model produced by weight-averaging a hyperparameter sweep, a technique its authors called a model soup, reported 90.94% top-1 in 2022 [50], and CoCa's 91.0% dates from the same year [36]. Google's 22-billion-parameter ViT demonstrated that transformer scaling works in vision but was presented mainly as a recipe and an analysis of scale effects rather than a leaderboard entry [51]. The DINOv3 authors' own comparison of frozen encoders under a common linear-probe protocol clusters tightly: 89.3% for Meta's Perception Encoder, 89.1% for SigLIP 2, 88.4% for DINOv3 itself and 87.9% for Apple's AIMv2, with ViT-22B listed at 89.5% under a different evaluation protocol [39]. Apple reports 89.5% for AIMv2-3B on its own frozen-trunk protocol, which illustrates how much these figures depend on the evaluation recipe [53]. Meta's Perception Encoder paper leads instead with an average zero-shot ImageNet robustness score of 86.6 [52].

The DINOv3 authors describe self-supervised learning as having "reached the Imagenet accuracy plateau of recent years" and argue that the interesting frontier is now dense features for segmentation and depth rather than global classification accuracy [39]. Practical work has shifted accordingly: fine-tuning or probing a general-purpose vision encoder on a domain-specific label set, rather than training a classifier from scratch, and using vision-language models to classify by prompt in settings where the label vocabulary is not known in advance.

## See also

- [ImageNet](https://aiwiki.ai/wiki/imagenet)
- [Convolutional neural network](https://aiwiki.ai/wiki/convolutional_neural_network)
- [AlexNet](https://aiwiki.ai/wiki/alexnet)
- [ResNet](https://aiwiki.ai/wiki/resnet)
- [Vision transformer](https://aiwiki.ai/wiki/vision_transformer)
- [Transfer learning](https://aiwiki.ai/wiki/transfer_learning)

## References

1. Russakovsky, O. et al. "ImageNet Large Scale Visual Recognition Challenge." arXiv:1409.0575 (2014, rev. 2015). https://arxiv.org/abs/1409.0575
2. ImageNet. "Large Scale Visual Recognition Challenge 2012 (ILSVRC2012) Results." https://image-net.org/challenges/LSVRC/2012/results
3. Krizhevsky, A., Sutskever, I., Hinton, G. E. "ImageNet Classification with Deep Convolutional Neural Networks." Advances in Neural Information Processing Systems 25 (2012). https://proceedings.neurips.cc/paper_files/paper/2012/file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf
4. ImageNet. "ImageNet Large Scale Visual Recognition Challenge (ILSVRC)" index of challenge years. https://www.image-net.org/challenges/LSVRC/
5. Yosinski, J., Clune, J., Bengio, Y., Lipson, H. "How transferable are features in deep neural networks?" arXiv:1411.1792 (2014). https://arxiv.org/abs/1411.1792
6. Kornblith, S., Shlens, J., Le, Q. V. "Do Better ImageNet Models Transfer Better?" arXiv:1805.08974 (2018, rev. 2019). https://arxiv.org/abs/1805.08974
7. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P. "Gradient-Based Learning Applied to Document Recognition." Proceedings of the IEEE, November 1998. http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf
8. He, T. et al. "Bag of Tricks for Image Classification with Convolutional Neural Networks." arXiv:1812.01187 (2018). https://arxiv.org/abs/1812.01187
9. Wightman, R., Touvron, H., Jegou, H. "ResNet strikes back: An improved training procedure in timm." arXiv:2110.00476 (2021). https://arxiv.org/abs/2110.00476
10. Ioffe, S., Szegedy, C. "Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift." arXiv:1502.03167 (2015). https://arxiv.org/abs/1502.03167
11. Krizhevsky, A. "The CIFAR-10 and CIFAR-100 datasets." University of Toronto. https://www.cs.toronto.edu/~kriz/cifar.html
12. Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., Fei-Fei, L. "ImageNet: A Large-Scale Hierarchical Image Database." CVPR 2009. https://www.image-net.org/static_files/papers/imagenet_cvpr09.pdf
13. Yang, K., Qinami, K., Fei-Fei, L., Deng, J., Russakovsky, O. "Towards Fairer Datasets" research post, ImageNet, 17 September 2019. https://www.image-net.org/update-sep-17-2019.php
14. Ciresan, D., Meier, U., Schmidhuber, J. "Multi-column Deep Neural Networks for Image Classification." arXiv:1202.2745 (2012). https://arxiv.org/abs/1202.2745
15. ImageNet. "Large Scale Visual Recognition Challenge 2010 (ILSVRC2010) Results." https://image-net.org/challenges/LSVRC/2010/results
16. ImageNet. "Large Scale Visual Recognition Challenge 2011 (ILSVRC2011) Results." https://image-net.org/challenges/LSVRC/2011/results
17. ImageNet. "ILSVRC2013 Results." https://image-net.org/challenges/LSVRC/2013/results
18. Zeiler, M. D., Fergus, R. "Visualizing and Understanding Convolutional Networks." arXiv:1311.2901 (2013). https://arxiv.org/abs/1311.2901
19. Szegedy, C. et al. "Going Deeper with Convolutions." arXiv:1409.4842 (2014). https://arxiv.org/abs/1409.4842
20. ImageNet. "ILSVRC2014 Results." https://image-net.org/challenges/LSVRC/2014/results
21. He, K., Zhang, X., Ren, S., Sun, J. "Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification." arXiv:1502.01852 (2015). https://arxiv.org/abs/1502.01852
22. He, K., Zhang, X., Ren, S., Sun, J. "Deep Residual Learning for Image Recognition." arXiv:1512.03385 (2015). https://arxiv.org/abs/1512.03385
23. ImageNet. "ILSVRC2016 Results." https://image-net.org/challenges/LSVRC/2016/results
24. ImageNet. "ILSVRC2017 Results." https://image-net.org/challenges/LSVRC/2017/results
25. Hu, J., Shen, L., Albanie, S., Sun, G., Wu, E. "Squeeze-and-Excitation Networks." arXiv:1709.01507 (2017, rev. 2019). https://arxiv.org/abs/1709.01507
26. Simonyan, K., Zisserman, A. "Very Deep Convolutional Networks for Large-Scale Image Recognition." arXiv:1409.1556 (2014, rev. 2015). https://arxiv.org/abs/1409.1556
27. Huang, G., Liu, Z., van der Maaten, L., Weinberger, K. Q. "Densely Connected Convolutional Networks." arXiv:1608.06993 (2016, rev. 2018). https://arxiv.org/abs/1608.06993
28. Howard, A. G. et al. "MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications." arXiv:1704.04861 (2017). https://arxiv.org/abs/1704.04861
29. Tan, M., Le, Q. V. "EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks." arXiv:1905.11946 (2019, rev. 2020). https://arxiv.org/abs/1905.11946
30. Xie, Q., Luong, M.-T., Hovy, E., Le, Q. V. "Self-training with Noisy Student improves ImageNet classification." arXiv:1911.04252 (2019, rev. 2020). https://arxiv.org/abs/1911.04252
31. Ridnik, T., Ben-Baruch, E., Noy, A., Zelnik-Manor, L. "ImageNet-21K Pretraining for the Masses." arXiv:2104.10972 (2021). https://arxiv.org/abs/2104.10972
32. Dosovitskiy, A. et al. "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale." arXiv:2010.11929 (2020, rev. 2021). https://arxiv.org/abs/2010.11929
33. Liu, Z. et al. "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows." arXiv:2103.14030 (2021). https://arxiv.org/abs/2103.14030
34. Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., Xie, S. "A ConvNet for the 2020s." arXiv:2201.03545 (2022). https://arxiv.org/abs/2201.03545
35. Radford, A. et al. "Learning Transferable Visual Models From Natural Language Supervision." arXiv:2103.00020 (2021). https://arxiv.org/abs/2103.00020
36. Yu, J. et al. "CoCa: Contrastive Captioners are Image-Text Foundation Models." arXiv:2205.01917 (2022). https://arxiv.org/abs/2205.01917
37. Tschannen, M. et al. "SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features." arXiv:2502.14786 (2025). https://arxiv.org/abs/2502.14786
38. Oquab, M. et al. "DINOv2: Learning Robust Visual Features without Supervision." arXiv:2304.07193 (2023, rev. 2024). https://arxiv.org/abs/2304.07193
39. Simeoni, O. et al. "DINOv3." arXiv:2508.10104 (2025). https://arxiv.org/abs/2508.10104
40. Recht, B., Roelofs, R., Schmidt, L., Shankar, V. "Do ImageNet Classifiers Generalize to ImageNet?" arXiv:1902.10811 (2019). https://arxiv.org/abs/1902.10811
41. Beyer, L., Henaff, O. J., Kolesnikov, A., Zhai, X., van den Oord, A. "Are we done with ImageNet?" arXiv:2006.07159 (2020). https://arxiv.org/abs/2006.07159
42. Hendrycks, D., Dietterich, T. "Benchmarking Neural Network Robustness to Common Corruptions and Perturbations." arXiv:1903.12261 (2019). https://arxiv.org/abs/1903.12261
43. Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., Song, D. "Natural Adversarial Examples." arXiv:1907.07174 (2019, rev. 2021). https://arxiv.org/abs/1907.07174
44. Geirhos, R. et al. "ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness." arXiv:1811.12231 (2018, rev. 2022). https://arxiv.org/abs/1811.12231
45. Szegedy, C. et al. "Intriguing properties of neural networks." arXiv:1312.6199 (2013, rev. 2014). https://arxiv.org/abs/1312.6199
46. Yang, K., Qinami, K., Fei-Fei, L., Deng, J., Russakovsky, O. "Towards Fairer Datasets: Filtering and Balancing the Distribution of the People Subtree in the ImageNet Hierarchy." arXiv:1912.07726 (2019). https://arxiv.org/abs/1912.07726
47. Esteva, A. et al. "Dermatologist-level classification of skin cancer with deep neural networks." Nature 542(7639), 2 February 2017. https://pubmed.ncbi.nlm.nih.gov/28117445/
48. Gulshan, V. et al. "Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs." JAMA 316(22), 13 December 2016. https://pubmed.ncbi.nlm.nih.gov/27898976/
49. PyTorch. "Models and pre-trained weights", torchvision documentation. https://pytorch.org/vision/main/models.html
50. Wortsman, M. et al. "Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time." arXiv:2203.05482 (2022). https://arxiv.org/abs/2203.05482
51. Dehghani, M. et al. "Scaling Vision Transformers to 22 Billion Parameters." arXiv:2302.05442 (2023). https://arxiv.org/abs/2302.05442
52. Bolya, D. et al. "Perception Encoder: The best visual embeddings are not at the output of the network." arXiv:2504.13181 (2025). https://arxiv.org/abs/2504.13181
53. Fini, E. et al. "Multimodal Autoregressive Pre-training of Large Vision Encoders." arXiv:2411.14402 (2024). https://arxiv.org/abs/2411.14402