Label smoothing
Label smoothing is a regularization technique for training classifiers in which the one-hot target label is replaced by a mixture of that label and a fixed distribution over classes, usually the uniform distribution, so that the model is trained to assign slightly less than full probability to the correct class.[1] It was introduced by Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens and Zbigniew Wojna in "Rethinking the Inception Architecture for Computer Vision" (CVPR 2016), the paper that defined Inception-v3, where the authors called it label-smoothing regularization (LSR).[1] The method changes only the training targets and adds one hyperparameter, the smoothing factor $\varepsilon$; the original paper, the Transformer and several later training recipes used 0.1.[1][4][5]
Label smoothing was adopted in many image classification training recipes and was used in the original Transformer for machine translation.[2][4][15] Research since 2019 has found that its effects go beyond a small accuracy gain. A study by Rafael Müller, Simon Kornblith and Geoffrey Hinton showed that it improves calibration but makes a network a worse teacher for knowledge distillation, and later work linked it to weaker transfer features, robustness to label noise, degraded selective classification and changed privacy risk.[2][10][13][15][17]
Definition
Hard targets and the cross-entropy gradient
A classifier with $K$ classes produces logits $z_1, \dots, z_K$, which the softmax turns into probabilities $p(k) = \exp(z_k) / \sum_{i=1}^{K} \exp(z_i)$. Training minimizes the cross-entropy $\ell = -\sum_{k=1}^{K} q(k) \log p(k)$ between a target distribution $q$ and the prediction $p$. The gradient with respect to each logit is $\partial \ell / \partial z_k = p(k) - q(k)$, which is bounded between -1 and 1.[1]
With a single correct label $y$, the target is the one-hot vector $q(k) = \delta_{k,y}$ (1 for the correct class, 0 otherwise). Szegedy and colleagues noted that this target cannot be reached with finite logits; it is only approached as the correct logit grows much larger than all the others. They identified two problems with that behavior. First, it can cause overfitting, because assigning full probability to every training label does not guarantee generalization. Second, it pushes the gap between the largest logit and the rest to grow, which, combined with the bounded gradient, reduces the model's ability to adapt. In the authors' words, the model "becomes too confident about its predictions".[1]
The smoothed target
Label smoothing replaces the one-hot target with
where $u(k)$ is a fixed distribution over labels that does not depend on the training example and $\varepsilon$ is the smoothing parameter.[1] Szegedy and colleagues suggested using the prior distribution over labels for $u$, but their experiments used the uniform distribution $u(k) = 1/K$, giving
so the correct class receives a target of $1 - \varepsilon + \varepsilon/K$ and every other class receives $\varepsilon/K$.[1] Because every entry of $q'$ now has a positive lower bound, a prediction that puts nearly all of its mass on one class incurs a large cross-entropy, and the largest logit is discouraged from becoming much larger than the others.[1]
The paper gives a sampling interpretation: the smoothed target is the distribution of a label produced by starting with the true label $y$ and, with probability $\varepsilon$, replacing it with a sample from $u$. The authors therefore described LSR as estimating "the marginalized effect of label-dropout during training".[1]
Equivalent forms of the loss
The cross-entropy with the smoothed target splits into two terms:[1]
The first term is the ordinary loss on the hard label. The second penalizes the deviation of the prediction $p$ from the prior $u$. Since $H(u, p) = D_{\mathrm{KL}}(u ,|, p) + H(u)$ and $H(u)$ is a constant, the second term is equivalent to a KL divergence penalty pulling the prediction toward the uniform distribution.[1] Gabriel Pereyra and colleagues at Google Brain pointed out in 2017 that reversing the direction of this divergence, penalizing $D_{\mathrm{KL}}(p ,|, u)$ instead, yields a "confidence penalty" that penalizes low-entropy output distributions.[3]
Common variants
| Variant | Target for true class | Target for each other class | Example sources |
|---|---|---|---|
| Uniform (original) | $1 - \varepsilon + \varepsilon/K$ | $\varepsilon/K$ | Szegedy et al.; PyTorch; Keras [1][7][9] |
| Uniform over incorrect classes | $1 - \varepsilon$ | $\varepsilon/(K-1)$ | He et al. (Bag of Tricks); DeiT [5][6] |
| Prior or unigram | depends on label frequencies | depends on label frequencies | Szegedy et al. (proposed); Pereyra et al. [1][3] |
| Online label smoothing | learned from model predictions | learned from model predictions | Zhang et al. [14] |
In the second variant, the mass removed from the true class is spread only over the incorrect classes. Tong He and colleagues showed that for this form the optimal logits leave a finite gap of $\log((K-1)(1-\varepsilon)/\varepsilon)$ between the correct class and the others, which shrinks as $\varepsilon$ grows; with $\varepsilon = 0.1$ and $K = 1000$ the gap is about 9.1. In their measurements on the ImageNet validation set, a ResNet-50-D trained with label smoothing had a distribution of gaps centred near that value with fewer extreme values than one trained without it.[5]
History
Inception-v3
Label smoothing appeared in section 7 of the Inception paper, "Model Regularization via Label Smoothing". On ILSVRC 2012 with $K = 1000$ classes, the authors used $u(k) = 1/1000$ and $\varepsilon = 0.1$ and reported "a consistent improvement of about 0.2% absolute" in both top-1 and top-5 error.[1] In the paper's table of cumulative changes, adding label smoothing to the Inception-v2 baseline trained with RMSProp lowered single-crop top-1 error from 23.1% to 22.8% and top-5 error from 6.3% to 6.1%. The final network with all of the paper's changes, including label smoothing, is what the authors called Inception-v3.[1]
Spread to other tasks
Pereyra and colleagues evaluated label smoothing and their confidence penalty on six benchmarks covering image classification (MNIST and CIFAR-10), language modeling (Penn Treebank), machine translation (WMT'14 English-to-German) and speech recognition (TIMIT and WSJ). They reported that both regularizers improved state-of-the-art models "without modifying existing hyperparameters". On MNIST and CIFAR-10 they searched smoothing values from 0.05 to 0.5 and found 0.1 worked best.[3]
The Transformer paper, "Attention Is All You Need" (2017), trained with label smoothing of $\varepsilon_{ls} = 0.1$ and noted: "This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score."[4] Müller and colleagues later summarized published results in which label smoothing raised English-to-German BLEU for the Transformer from 25.3 to 25.8 while perplexity rose from 4.67 to 4.92, and in which Jan Chorowski and Navdeep Jaitly's attention-based speech recognizer on WSJ improved its word error rate from 8.9 to 7.0/6.7.[2]
Standard training recipes
The 2019 "Bag of Tricks" study from Amazon Web Services listed label smoothing among four training refinements for convolutional neural networks. Following Szegedy and colleagues it set $\varepsilon = 0.1$; adding label smoothing to a ResNet-50-D already trained with cosine learning-rate decay raised ImageNet top-1 accuracy from 77.91% to 78.31%.[5] The DeiT recipe for training vision transformers on ImageNet (DeiT, 2020) fixed $\varepsilon = 0.1$ in all of its experiments that use true labels.[6]
Major deep learning frameworks expose label smoothing as a loss argument. PyTorch added a label_smoothing argument to nn.CrossEntropyLoss in version 1.10, released in October 2021; the documentation describes the targets as "a mixture of the original ground truth and a uniform distribution" as in the Inception paper, with a default of 0.0.[7][8] In TensorFlow and Keras, CategoricalCrossentropy takes a label_smoothing value in [0, 1]; with 0.1, non-target labels become 0.1/num_classes and the target becomes 0.9 + 0.1/num_classes.[9]
Effects on representations and calibration
Tighter clusters in the penultimate layer
Müller, Kornblith and Hinton's NeurIPS 2019 paper "When Does Label Smoothing Help?" set out to explain why and when the technique works. The authors argued that label smoothing pushes the difference between the correct logit and each incorrect logit toward a constant that depends on $\varepsilon$. Treating each logit as a measure of distance between the penultimate-layer activations and a per-class template, this means each training example is encouraged to sit close to its own class template and equally far from the templates of all other classes.[2]
To show this, they projected penultimate-layer activations onto the plane through the templates of three classes. For AlexNet on CIFAR-10, ResNet-56 on CIFAR-100 and Inception-v4 on ImageNet, networks trained with $\varepsilon = 0.1$ formed much tighter clusters arranged in regular triangles, while networks trained with hard targets formed broader clusters. For two similar ImageNet classes (toy poodle and miniature poodle) shown alongside an unrelated class (tench), hard-target training preserved a continuous gradation between clusters that, in the authors' phrase, could measure "how much a poodle is a particular tench"; with label smoothing that information was "virtually erased".[2] Accuracy changed little in these examples: 86.8% versus 86.7% for AlexNet on CIFAR-10, 72.1% versus 72.7% for ResNet-56 on CIFAR-100, and 80.9% for Inception-v4 on ImageNet in both cases.[2]
Implicit calibration
A classifier is well calibrated when its predicted confidence matches its accuracy, commonly measured with expected calibration error (ECE). Müller and colleagues found that label smoothing reduces ECE to a level similar to post-hoc temperature scaling, without a separate calibration step.[2]
| Setting | ECE, hard targets | ECE, temperature scaling | ECE, label smoothing |
|---|---|---|---|
| ResNet-56, CIFAR-100 | 0.150 | 0.021 (T = 1.9) | 0.024 ($\varepsilon$ = 0.05) |
| Inception-v4, ImageNet | 0.071 | 0.022 (T = 1.4) | 0.035 ($\varepsilon$ = 0.1) |
| Transformer, English-German | 0.056 | 0.018 (T = 1.13) | 0.019 ($\varepsilon$ = 0.1) |
Source: Müller, Kornblith and Hinton (2019), table 3.[2]
For translation, the authors found that the label-smoothed Transformer was better calibrated and scored higher BLEU but had worse negative log-likelihood than the hard-target model at every temperature. Temperature scaling of the hard-target model reduced its calibration error and raised BLEU slightly, but not enough to match the smoothed model. They concluded that label smoothing's BLEU advantage after beam search is "only partly explained by calibration".[2]
Effect on transfer learning
Simon Kornblith, Ting Chen, Honglak Lee and Mohammad Norouzi compared nine training objectives for ResNet-50 on ImageNet (NeurIPS 2021). Label smoothing raised ImageNet top-1 accuracy from 77.0% (softmax cross-entropy) to 77.6%, but its fixed penultimate-layer features transferred worse to most downstream datasets under linear evaluation; on the Cars dataset, for example, linear transfer accuracy fell from 60.3% to 54.8%.[13] The authors found that objectives producing greater class separation in the penultimate layer gave higher ImageNet accuracy but less useful features for other tasks, and that the choice of loss had little effect when networks were fully fine-tuned on the new task.[13] Müller and colleagues had already cited earlier evidence that label smoothing can hurt transfer learning.[2]
Relationship to knowledge distillation
Müller et al.: teachers trained with label smoothing
Knowledge distillation also trains a model on soft targets, in this case the temperature-scaled outputs of a teacher network. Müller and colleagues found that teachers trained with label smoothing produce worse students, even when smoothing makes the teacher more accurate.[2] They first noticed the effect while reproducing an MNIST experiment: a teacher trained with dropout reached 0.67% test error and distilled a student to 0.74%, while a teacher trained with label smoothing instead trained faster and reached 0.59%, but its student reached only 0.91%.[2]
In a CIFAR-10 experiment distilling a ResNet-56 teacher into an AlexNet student, increasing $\varepsilon$ improved the teacher's accuracy up to $\varepsilon = 0.6$, yet students of these better teachers did no better, and sometimes worse, than a student trained directly with label smoothing. Students distilled from hard-target teachers outperformed that baseline.[2] The authors attributed the gap to "erasure" of information: their estimate of the mutual information between training examples and the difference of two logits rose early in training and then slowly fell, especially for the label-smoothed teacher, which they took as evidence that information distinguishing individual examples is lost.[2]
Later challenges and reinterpretations
Li Yuan and colleagues (CVPR 2020) argued in the other direction that knowledge distillation is itself "a type of learned label smoothing regularization" and that label smoothing provides a "virtual teacher model" for distillation. Their Teacher-free Knowledge Distillation framework, in which a student learns from itself or a manually designed distribution, achieved up to 0.65% improvement on ImageNet over baseline models according to the authors.[11]
Zhiqiang Shen and colleagues (ICLR 2021) tested the incompatibility claim on ImageNet, fine-grained CUB200-2011, binary neural networks and neural machine translation. They reported that the erasing effect harms mainly semantically different classes, which are already easy to separate, while for semantically similar classes it pushes the clusters further apart. In their experiments the teacher's accuracy was the dominant factor in student quality whether or not the teacher used label smoothing, which they described as "partially against" Müller and colleagues' conclusion, and in machine translation a stronger label-smoothed teacher still distilled a higher-BLEU student.[12]
Michal Lukasik and colleagues at Google Research (ICML 2020) found a related exception: when distilling from data with noisy labels, label smoothing of the teacher was beneficial.[10]
Label noise and robustness
Label smoothing can be read as injecting uniform noise into every label, which suggests it might make learning from mislabeled data harder.[10] Lukasik and colleagues showed that it is closely related to loss-correction techniques from the label-noise literature and found empirically that it is competitive with loss correction under label noise and has "strong denoising effects". They explained the effect by relating smoothing to L2 regularization.[10] Chang-Bin Zhang and colleagues proposed Online Label Smoothing, which builds the soft target for each class from the statistics of the model's own predictions instead of a fixed uniform distribution; they reported gains on CIFAR-100, ImageNet and fine-grained datasets and better robustness to noisy labels than standard label smoothing.[14]
Ali Shafahi and colleagues at the University of Maryland reported in 2019 that label smoothing and logit squeezing, combined with Gaussian noise injection, could produce strong robustness to adversarial attacks without generating any adversarial examples, often exceeding that of adversarial training in their experiments.[16] Later analysis questioned how much of that robustness is real: Chaohao Fu and colleagues found that a label-smoothed WideResNet could reach up to 75% accuracy under PGD attack on CIFAR-10, but described this robustness as "incomplete", because the defense effect was volatile and did not hold against attacks transferred from a naturally trained model.[21]
Limitations
Several studies have identified settings where label smoothing does not help or causes harm:
| Finding | Setting | Source |
|---|---|---|
| Worse students when used on the teacher | Knowledge distillation on clean data (MNIST, CIFAR-10) | Müller et al. 2019 [2] |
| Less transferable fixed features | Linear and k-nearest-neighbor transfer from ImageNet | Kornblith et al. 2021 [13] |
| Lost or reduced benefit | Long-tailed class distributions and larger numbers of classes | Shen et al. 2021 [12] |
| Degraded selective classification | Rejecting likely errors using the model's confidence | Xia et al. 2025 [15] |
| Increased model inversion leakage | Model inversion attacks on classifiers | Struppek et al. 2024 [17] |
| Diminished calibration benefit | Supervised fine-tuning of large-vocabulary LLMs | Huang et al. 2025 [19] |
Shen and colleagues observed that long-tailed datasets (ImageNet-LT, Places365-LT and iNaturalist 2019) suppress the effectiveness of label smoothing, and that its improvement generally shrinks as the number of classes grows.[12]
Guoxuan Xia and colleagues (ICLR 2025) showed across a range of large-scale tasks and architectures that label smoothing consistently degrades selective classification, in which a model abstains when uncertain. Their gradient analysis found that label smoothing suppresses the maximum logit more when a prediction is likely correct and less when it is likely wrong, which weakens the ranking of correct versus incorrect predictions by confidence. They showed that post-hoc logit normalization can recover selective classification performance lost to label smoothing.[15]
Lukas Struppek, Dominik Hintersdorf and Kristian Kersting (ICLR 2024) found that conventional label smoothing makes classifiers more vulnerable to model inversion attacks, which reconstruct class-representative samples and can leak information about training data. Smoothing with negative factors had the opposite effect and acted as a privacy defense in their experiments.[17]
Later research
Research on label smoothing continues, much of it on why the technique works and on variants that address its side effects. Examples published in 2025 include:
| Year | Work | Contribution |
|---|---|---|
| 2025 | Guo et al., TMLR | Analyzes label smoothing through neural collapse; models trained with it converge faster to neural-collapse solutions [18] |
| 2025 | Huang, Lu and Zeng, ICML | Finds calibration degrades after instruction tuning of open LLMs and studies label smoothing during supervised fine-tuning [19] |
| 2025 | Zhou et al. (MaxSup), NeurIPS | Penalizes the top-1 logit instead of the ground-truth logit to avoid overconfidence on misclassified samples [20] |
Li Guo and colleagues reported that models trained with label smoothing reach a stronger level of neural collapse, a framework that describes model behavior in the terminal phase of training, and derived closed-form minimizers suggesting that label smoothing converges faster in theory.[18] Jerry Huang, Peng Lu and Qiuhao Zeng examined open-source large language models and found significant calibration degradation after instruction tuning. They explained why label smoothing can maintain calibration during supervised fine-tuning, but found settings where its effect is severely diminished, which they traced to the hidden and vocabulary sizes of the models, and they wrote a custom kernel to cut the memory cost of the label-smoothed cross-entropy loss.[19] Yuxuan Zhou and colleagues decomposed the label smoothing loss into a regularization term that acts only on correct predictions and an "error-amplification term" that, they argued, reinforces incorrect predictions and contributes to representation collapse; their Max Suppression (MaxSup) method applies the penalty to the top-1 logit instead.[20]
Continue reading
- DropoutDropout is a regularization technique for neural networks that randomly sets a fraction of neuron activations to zero during trainingMentioned twice above
- One-Hot EncodingOne-hot encoding is a data preprocessing technique that converts a categorical variable with…Listed in See also
- Calibration LayerA calibration layer is a post-prediction adjustment appended to a trained machine learning model that rescales its raw output scores or predicted probabilities so they better reflect the true likelihood of…Listed in See also
- Focal lossFocal loss is a loss function that reshapes standard cross-entropy loss by adding a (1 - p_t)^gamma modulating factor, which down-weights well-classified (easy) examples so that training concentrates on hardListed in See also
- Data AugmentationData augmentation is a set of techniques that artificially expand the size and diversity of a training dataset by applying label-preserving transformations to existing examples, rather than collecting new…Listed in See also
- Confident Learning (CL)Confident Learning (CL) is a data-centric machine learning framework for characterizing, finding, and learning with label errors in datasets.Listed in See also
More to read
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z. "Rethinking the Inception Architecture for Computer Vision". Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 2818-2826. openaccess.thecvf.com/...Inception_CVPR_2016_paper
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Müller, R., Kornblith, S., Hinton, G. "When Does Label Smoothing Help?". Advances in Neural Information Processing Systems 32 (NeurIPS 2019). arxiv.org/...1906.02629
- ^1 ^2 ^3Pereyra, G., Tucker, G., Chorowski, J., Kaiser, Ł., Hinton, G. "Regularizing Neural Networks by Penalizing Confident Output Distributions". arXiv:1701.06548, January 2017. arxiv.org/...1701.06548
- ^1 ^2 ^3Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., Polosukhin, I. "Attention Is All You Need". 31st Conference on Neural Information Processing Systems (NIPS 2017). arxiv.org/...1706.03762
- ^1 ^2 ^3 ^4He, T., Zhang, Z., Zhang, H., Zhang, Z., Xie, J., Li, M. "Bag of Tricks for Image Classification with Convolutional Neural Networks". Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 558-567. openaccess.thecvf.com/..._Networks_CVPR_2019_paper
- ^1 ^2Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H. "Training data-efficient image transformers & distillation through attention". Proceedings of the 38th International Conference on Machine Learning (ICML 2021), PMLR 139, pp. 10347-10357. arxiv.org/...2012.12877
- ^1 ^2PyTorch. "CrossEntropyLoss". PyTorch 2.15 documentation. docs.pytorch.org/...torch.nn.CrossEntropyLoss
- ^PyTorch. "PyTorch 1.10 Release, including CUDA Graphs APIs, Frontend and compiler improvements". GitHub release notes, 2021-10-21. github.com/...v1.10.0
- ^1 ^2TensorFlow. "tf.keras.losses.CategoricalCrossentropy". TensorFlow API documentation. tensorflow.org/...CategoricalCrossentropy
- ^1 ^2 ^3 ^4Lukasik, M., Bhojanapalli, S., Menon, A., Kumar, S. "Does label smoothing mitigate label noise?". Proceedings of the 37th International Conference on Machine Learning (ICML 2020), PMLR 119, pp. 6448-6458. proceedings.mlr.press/...lukasik20a
- ^Yuan, L., Tay, F. E. H., Li, G., Wang, T., Feng, J. "Revisiting Knowledge Distillation via Label Smoothing Regularization". Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 3903-3911. openaccess.thecvf.com/...arization_CVPR_2020_paper
- ^1 ^2 ^3Shen, Z., Liu, Z., Xu, D., Chen, Z., Cheng, K.-T., Savvides, M. "Is Label Smoothing Truly Incompatible with Knowledge Distillation: An Empirical Study". International Conference on Learning Representations (ICLR 2021). arxiv.org/...2104.00676
- ^1 ^2 ^3 ^4Kornblith, S., Chen, T., Lee, H., Norouzi, M. "Why Do Better Loss Functions Lead to Less Transferable Features?". 35th Conference on Neural Information Processing Systems (NeurIPS 2021). arxiv.org/...2010.16402
- ^1 ^2Zhang, C.-B., Jiang, P.-T., Hou, Q., Wei, Y., Han, Q., Li, Z., Cheng, M.-M. "Delving Deep into Label Smoothing". IEEE Transactions on Image Processing, vol. 30, pp. 5984-5996, 2021. arxiv.org/...2011.12562
- ^1 ^2 ^3 ^4Xia, G., Laurent, O., Franchi, G., Bouganis, C.-S. "Towards Understanding Why Label Smoothing Degrades Selective Classification and How to Fix It". International Conference on Learning Representations (ICLR 2025). arxiv.org/...2403.14715
- ^Shafahi, A., Ghiasi, A., Huang, F., Goldstein, T. "Label Smoothing and Logit Squeezing: A Replacement for Adversarial Training?". arXiv:1910.11585, October 2019. arxiv.org/...1910.11585
- ^1 ^2 ^3Struppek, L., Hintersdorf, D., Kersting, K. "Be Careful What You Smooth For: Label Smoothing Can Be a Privacy Shield but Also a Catalyst for Model Inversion Attacks". International Conference on Learning Representations (ICLR 2024). arxiv.org/...2310.06549
- ^1 ^2Guo, L., Andriopoulos, G., Zhao, Z., Ling, S., Dong, Z., Ross, K. "Cross Entropy versus Label Smoothing: A Neural Collapse Perspective". Transactions on Machine Learning Research, May 2025. arxiv.org/...2402.03979
- ^1 ^2 ^3Huang, J., Lu, P., Zeng, Q. "Calibrated Language Models and How to Find Them with Label Smoothing". Forty-second International Conference on Machine Learning (ICML 2025). arxiv.org/...2508.00264
- ^1 ^2Zhou, Y., Li, H., Cheng, Z.-Q., Yan, X., Dong, Y., Fritz, M., Keuper, M. "MaxSup: Overcoming Representation Collapse in Label Smoothing". NeurIPS 2025. arxiv.org/...2502.15798
- ^Fu, C., Chen, H., Ruan, N., Jia, W. "Label Smoothing and Adversarial Robustness". arXiv:2009.08233, September 2020. arxiv.org/...2009.08233
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
2 revisions · v2 · 3,618 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check before publication on November 18, 2026: 78 claims checked against 26 cited and independent sources; 1 material and 2 minor corrections applied before publication.
Cite this page: AI Wiki. "Label smoothing." aiwiki.ai, updated 18 Nov 2026, fact-checked 18 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/label_smoothing


