Understanding deep learning (still) requires rethinking generalization
Open accessThe DOI matches its Crossref record; 45 of the 47 references it cites were found in a registry or an index; none of the 12 sources its claims cite is checked for support yet.
Abstract
Despite their massive size, successful deep artificial neural networks can exhibit a remarkably small gap between training and test performance. Conventional wisdom attributes small generalization error either to properties of the model family or to the regularization techniques used during training. Through extensive systematic experiments, we show how these traditional approaches fail to explain why large neural networks generalize well in practice. Specifically, our experiments establish that state-of-the-art convolutional networks for image classification trained with stochastic gradient methods easily fit a random labeling of the training data. This phenomenon is qualitatively unaffected by explicit regularization and occurs even if we replace the true images by completely unstructured random noise. We corroborate these experimental findings with a theoretical construction showing that simple depth two neural networks already have perfect finite sample expressivity as soon as the number of parameters exceeds the number of data points as it usually does in practice. We interpret our experimental findings by comparison with traditional models. We supplement this republication with a new section at the end summarizing recent progresses in the field since the original version of this paper.
Overview
CoChat Summary
TL;DR
Deep neural networks can fit random labels and random pixels with zero training error, so traditional complexity and regularization explanations do not account for their generalization. The paper argues that explicit regularization is neither necessary nor sufficient, and offers a finite-sample expressivity construction plus linear-model implicit-regularization analysis.
Key findings
- On CIFAR10, Inception, AlexNet, and MLPs reach 100% (or 99.82%/99.34%) training accuracy on random labels while test accuracy is near chance: Inception 9.78%, Inception w/o BatchNorm 10.12%, AlexNet 9.86%, MLP 3×512 10.48%, MLP 1×512 10.61%.
- On ImageNet, Inception reaches 95.20% top-1 accuracy on 1.2 million random labels from 1000 categories without hyperparameter tuning; with explicit regularizers on it still reaches ∼90% top-1.
- Explicit regularization may improve generalization but is neither necessary nor sufficient: on CIFAR10 all models still generalize well with regularizers off; on ImageNet top-1 accuracy without regularization is 59.80% vs 0.1% random guessing, and with data augmentation only Inception reaches 72.95% top-1.
- Theoretically, a two-layer ReLU network with p = 2n + d parameters can represent any labeling of any sample of size n in d dimensions; depth-k networks can use O(n/k) parameters per layer. In linear models, SGD converges to a minimum-norm solution, but minimum norm is not predictive: on MNIST the minimum norm with no preprocessing is ∼220 and with wavelet preprocessing ∼390, yet test error drops by a factor of 2.
Methods
- Randomization tests on CIFAR10 and ImageNet ILSVRC 2012: train Inception V3 on ImageNet and smaller Inception, AlexNet, and MLPs on CIFAR10 under true labels, partially corrupted labels, random labels, shuffled pixels, random pixels, and Gaussian pixels, toggling data augmentation, weight decay, and dropout.
- Compare training/test accuracy and convergence time; supplement with a finite-sample expressivity theorem for two-layer ReLU networks and with linear/kernel experiments on MNIST and CIFAR10 (Gaussian kernels, random convolutional filters).
Limitations
- Stated: the paper offers no new theory of generalization; it says a full mathematical characterization remains challenging and that it cannot fully survey recent progress. It also states that minimum-norm intuition is only a small piece of the generalization story and that it does not explain why certain architectures generalize better.
- Evident: the empirical scope is limited to image classification benchmarks (CIFAR10, ImageNet) and a small set of architectures/regularizers, so the findings may not extend to other domains or model families. The theoretical construction uses a depth-2 network with large width, and the paper notes depth-k alternatives with O(n/k) parameters per layer.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
0 Valid · 7 Not valid · 12 Not checked · 85 Own results and interpretations
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In other words, by randomizing labels alone we can force the generalization error of a model to jump up considerably without changing the model, its size, hyperparameters, or the optimizer.”
§ Understanding Deep Learning (Still) Requires Rethinking Generalization > 1.1. The randomization test
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
None of the sources cited for this claim have been checked yet.
In the paper
“although it does not reach the perfect 100% top-1 accuracy, 95.20% accuracy is still very surprising for 1.2 million random labels from 1000 categories.”
§ Understanding Deep Learning (Still) Requires Rethinking Generalization > 2. EFFECTIVE CAPACITY OF NEURAL NETWORKS
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The networks fit the corrupted training set perfectly for all the cases.”
§ Understanding Deep Learning (Still) Requires Rethinking Generalization > Partially corrupted labels.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“On the CIFAR10 dataset, Alexnet and MLPs all converge to zero loss on the training set.”
§ Understanding Deep Learning (Still) Requires Rethinking Generalization > 2.1. Fitting random labels and pixels
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We corroborate these experimental findings with a theoretical construction showing that simple depth two neural networks already have perfect finite sample expressivity as soon as the number of parameters exceeds the number of data points as it usually does in practice.”
- Also stated in: § 1.3. Finite sample expressivity · plain
“More formally, we exhibit a very simple two-layer ReLU network with p = 2n + d parameters that can express any labeling of any sample of size n in d dimensions.”
§ Understanding Deep Learning (Still) Requires Rethinking Generalization > Abstract
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 1.3. Finite sample expressivity · plain
70 more claims
Create a free account to see all 104 claims and every source verdict.
- It may state two separate claims as one. 17
- Probably not a substantial claim. 15
- The quote only partly states this claim. 1
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Based on existing research(Saha et al., 2021; Zhang et al., 2021), the following conclusion can be drawn: the gradient updates of a linear layer always lie within the space spanned by its input.”
From Knowledge Decoupling via Orthogonal Projection for Lifelong Editing of Large Language Models (2025) · cites this paper as Zhang et al., 2021
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“If a model is successfully trained on a dataset with random labels, which has been shown to be possible with deep neural networks [294], then it has learned the data generation process by memorizing the random labels.”
From From Anecdotal Evidence to Quantitative Evaluation Methods: A Systematic Review on Evaluating Explainable AI (2023) · cites this paper as [294]
The check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“Existing works in privacy-preserving federated learning are mostly developed based on the wellknown privacy-preserving techniques, including: (1) homomorphic encryption (HE), such as Paillier [90], Elgamal [91] and Brakerski-Gentry-Vaikuntanathan cryptosystems [92]; (2) Secure Multiparty Computation (SMC), such as garbled circuits [93] and secret sharing [94]; and (3) differential privacy (DP) [95”
From Privacy and Robustness in Federated Learning: Attacks and Defenses (2024) · cites this paper as [95]
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Understanding Deep Learning (Still) Requires Rethinking Generalization > 2.1. Fitting random labels and pixels
“On the CIFAR10 dataset, Alexnet and MLPs all converge to zero loss on the training set.”
§ Understanding Deep Learning (Still) Requires Rethinking Generalization > Rademacher complexity and VC-dimension.
“(fitting random labels) No No 99.82 9.86 MLP 3 × 512 1,735,178 No Yes 100.0 53.35No No 100.0 52.39 (fitting random labels) No No 100.0 10.48 MLP 1 × 512 1,209,866 No Yes 99.80 50.39No No 100.0 50.51 (fitting random labels) No No 99.34 10.61”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Understanding Deep Learning (Still) Requires Rethinking Generalization > 1.1. The randomization test
“More precisely, when trained on a completely random labeling of the true data, neural networks achieve 0 training error.”
§ Understanding Deep Learning (Still) Requires Rethinking Generalization > 2. EFFECTIVE CAPACITY OF NEURAL NETWORKS
“although it does not reach the perfect 100% top-1 accuracy, 95.20% accuracy is still very surprising for 1.2 million random labels from 1000 categories.”
Citations
Cites
(2014) · Cited by 33,419
Wu, Y. (1999) · Technometrics · doi:10.1080/00401706.1999.10485951 · Cited by 27,221
Advances in neural information processing systems 7
Cios, K., & Shields, M. (1997) · Neurocomputing · doi:10.1016/s0925-2312(97)90018-7 · Cited by 22,059
Approximation by superpositions of a sigmoidal function
Cybenko, G. (1989) · Mathematics of Control, Signals, and Systems · doi:10.1007/bf02551274 · Cited by 13,883
Advances in Neural Information Processing Systems 25
Zhang, Y., Sutton, C. A., Storkey, A., & Zoubin Ghahramani, A. (2012) · Cited by 12,653
Understanding deep learning requires rethinking generalization
Zhang, C., Bengio, S., Hardt, M., Recht, B., & Vinyals, O. (2016) · arXiv · doi:10.48550/arxiv.1611.03530 · Cited by 2,838
Self-citationReconciling modern machine-learning practice and the classical bias–variance trade-off
Belkin, M., Hsu, D., Ma, S., & Mandal, S. (2019) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.1903070116 · Cited by 1,595 · cited 1×
“23, 7, 6, 16 Despite significant progress on theoretical understanding of deep learning in the past few years, a full mathematical characterization of the whole story remains challenging.”
BackgroundA Generalized Representer Theorem
Schölkopf, B., Herbrich, R., & Smola, A. J. (2001) · Lecture Notes in Computer Science · doi:10.1007/3-540-44581-1_27 · Cited by 1,589 · cited 1×
“We have thus derived the “kernel trick” 34 —albeit in a round-about fashion.”
MethodA Closer Look at Memorization in Deep Networks
Arpit, D., Jastrzębski, S., Ballas, N., Krueger, D., Bengio, E., Kanwal, M. S., Maharaj, T., Fischer, A., Courville, A., Bengio, Y., et al. (2017) · arXiv · doi:10.48550/arxiv.1706.05394 · Cited by 1,307
Bartlett, P. L. (1998) · IEEE Transactions on Information Theory · doi:10.1109/18.661502 · Cited by 1,166
Research works printed but not linked: 8
- Cohen, N., Shashua, A. Convolutional rectifier networks as generalized tensor decompositions. In International Conference on Machine Learning. M.F. Balcan and K.Q. Weinberger, eds. 2016, 955--963.
- Cybenko, G. Approximation by superposition of sigmoidal functions. Math. Control Signal. Syst. 4, 2 (1989), 303--314.
- Golowich, N., Rakhlin, A., Shamir, O. Size-independent sample complexity of neural networks. In Conference On Learning Theory. P.R. Sébastien Bubeck, V. Perchet, eds., 2018, 297--299.
- Ma, S., Bassily, R., Belkin, M. The power of interpolation: Understanding the effectiveness of sgd in modern over-parametrized learning. In International Conference on Machine Learning. J. Dy and A. Krause, eds. 2018, 3325--3334.
- Morcos, A., Raghu, M., Bengio, S. Insights on representational similarity in neural networks with canonical correlation. Adv. Neural Inform. Process. Syst.. 2018, 5727--5736.
- Neyshabur, B., Tomioka, R., Srebro, N. Norm-based capacity control in neural networks. In Conference on Learning Theory. S.K. Peter Grünwald and E. Hazan, eds. 2015, 1376--1401.
- Smith, S.L., Le, Q.V. A bayesian perspective on generalization and stochastic gradient descent. In International Conference on Learning Representations, 2018.
- Zhou, W., Veitch, V., Austern, M., Adams, R.P., Orbanz, P. Non-vacuous generalization bounds at the ImageNet scale: A PAC-Bayesian compression approach. In International Conference on Learning Representations, 2019.
Cited by
Items 1,201–1,300 of 2,667
Performance Analysis of QR Phishing Detection Approaches
Nigam, N. (2025) · Journal of Information Systems Engineering & Management · doi:10.52783/jisem.v10i33s.5472 · Cited by 5
“Zhang et al. [18] stress the need to enhance deep learning generalizability, aiding further CNN model tuning to adapt to diverse QR Code representations and email content.”
Khoza, R. T. (2025) · Springer proceedings in business and economics · doi:10.1007/978-3-031-84885-8_53 · Cited by 5
A Practical Theory of Generalization in Selectivity Learning
Wu, P., Xu, H., Marcus, R., & Ives, Z. G. (2025) · Proceedings of the VLDB Endowment · doi:10.14778/3725688.3725708 · Cited by 5
Unsupervised Network Anomaly Detection with Autoencoders and Traffic Images
Neri, M., & Baldoni, S. (2025) · European Signal Processing Conference (EUSIPCO) · doi:10.23919/eusipco63237.2025.11226720 · Cited by 5
“84 GMACs, thus being equally lightweight [20].”
Enhancing weakly supervised 3D medical image segmentation through probabilistic-aware learning
Jiang, R., Fan, Z., Wu, J., Zhu, L., Huang, X., Wang, T., Huang, H., & Xu, M. (2026) · Pattern Recognition · doi:10.1016/j.patcog.2026.113973 · Cited by 5
“To mitigate these challenges, researchers have explored various strategies such as data augmentation [8, 9, 10], transfer learning [11, 12], and domain adaptation [13, 14] to reduce reliance on extensive labeled data.”
Drill the Cork of Information Bottleneck by Inputting the Most Important Data
Peng, X., Zhang, J., Wang, F., & Li, L. (2021) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2021.3079112 · Cited by 6
PeerRank: Robust Learning to Rank With Peer Loss Over Noisy Labels
Wu, X., Liu, Q., Qin, J., & Yu, Y. H. (2022) · IEEE Access · doi:10.1109/access.2022.3142096 · Cited by 6
Meta-Learning for Decoding Neural Activity Data With Noisy Labels
Xu, D., & Chen, R. (2022) · Frontiers in Computational Neuroscience · doi:10.3389/fncom.2022.913617 · Cited by 5
“In a study to assess the impact of noisy labels (Zhang et al., 2021), a test error rate of 0.1 was achieved by applying the Inception model to the CIFAR10 dataset, and the test error rate increased to 0.4 when 20% of labels were corrupted.”
MethodInfluentialRegularization in Network Optimization via Trimmed Stochastic Gradient Descent With Noisy Label
Nakamura, K., Sohn, B.-S., Won, K., & Hong, B. (2022) · IEEE Access · doi:10.1109/access.2022.3171910 · Cited by 5
Gionanidis, E., Karvelis, P. S., Georgoulas, G., Stamos, K., & Garg, P. (2022) · IEEE Secure Development Conference (SecDev) · doi:10.1109/secdev53368.2022.00017 · Cited by 5
“This phenomenon is magnified especially in cases when the available data are limited and the network has many layers [24].”
BackgroundInfluentialCross-Modal Data Augmentation for Tasks of Different Modalities
Chen, D., Zhuang, Y., Shen, Z., Yang, C., Wang, G., Tang, S., & Yang, Y. (2022) · IEEE Transactions on Multimedia · doi:10.1109/tmm.2022.3228696 · Cited by 9
“For example, a learned model may describe random error or noise instead of the underlying data distribution [1], and it may exhibit good performance on the training data but fail drastically on the testing data.”
BackgroundUlman, H., Gütter, J., & Niebling, J. (2023) · Frontiers in Remote Sensing · doi:10.3389/frsen.2022.1100012 · Cited by 7
“However, label noise in training data can potentially impair the performance of DNNs by damaging a network’s generalization ability, as it was empirically shown that DNNs are able to overfit to completely random noise (Zhang et al., 2021).”
BackgroundSelf-label correction for image classification with noisy labels
Zhang, Y., Fan, L., Mi, S., & Bian, Y. (2023) · Pattern Analysis and Applications · doi:10.1007/s10044-023-01180-w · Cited by 5
“Studies have demonstrated that deep neural networks with over-parameterization can fit data samples with random labels [4, 5] and have a significant impact on performance.”
BackgroundRobust Remote Sensing Image Cross-Scene Classification Under Noisy Environment
Zhu, S., Wu, C., Du, B., & Zhang, L. (2023) · IEEE Transactions on Geoscience and Remote Sensing · doi:10.1109/tgrs.2023.3337652 · Cited by 5
“Zhang et al. [10] demonstrate that convolutional neural networks (CNNs) can easily fit with corrupted labels, which severely degrades the performance and generalization ability on test datasets.”
BackgroundEnhancing Model Agnostic Meta-Learning via Gradient Similarity Loss
Tak, J.-H., & Hong, B. (2024) · Electronics · doi:10.3390/electronics13030535 · Cited by 5
“However, when the amount of training data is limited, the risk of overfitting increases, particularly when the number of meta-parameters is relatively high [15].”
BackgroundUnder manipulations, are some AI models harder to audit?
Godinot, A., Merrer, E. L., Trédan, G., Penzo, C., & Taïani, F. (2024) · IEEE Conference on Secure and Trustworthy Machine Learning (SaTML) · doi:10.1109/satml59370.2024.00038 · Cited by 11
“Moreover, Definition 2 is the literal formalization of the notion of benign overfitting considered in [22] and [23]: models that can fit any labeling –even random– of the train set while still having a good test performance when evaluated on the target distribution.”
InfluentialBenign overfitting and adaptive nonparametric regression
Chhor, J., Sigalla, S., & Tsybakov, A. B. (2024) · Probability Theory and Related Fields · doi:10.1007/s00440-024-01278-0 · Cited by 5
Dual-consistency constraints network for noisy facial expression recognition
Xia, H., Su, C., Song, S., & Tan, Y. (2024) · Image and Vision Computing · doi:10.1016/j.imavis.2024.105141 · Cited by 5
Lehmler, S. J., Saif-ur-Rehman, M., Glasmachers, T., & Iossifidis, I. (2024) · Neurocomputing · doi:10.1016/j.neucom.2024.128473 · Cited by 5
Enhancing Accuracy in Deep Learning Using Random Matrix Theory
Berlyand, L. V., Sandier, É., Shmalo, Y., & Zhang, L. (2024) · Journal of Machine Learning · doi:10.4208/jml.231220 · Cited by 10
Potapov, A. S., & Rodionov, S. A. (2017) · Lecture notes in computer science · doi:10.1007/978-3-319-63703-7_10 · Cited by 5
Shoham, N., & Avron, H. (2023) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2023.3287042 · Cited by 5
“The phenomena of benign overfitting and double descent was first recognized in DNNs [1], and later discussed and analyzed in the context of linear models [1], [2], [3], [11], [12].”
BackgroundInfluentialStealing Your Data from Compressed Machine Learning Models
Xu, N., Liu, Q., Liu, T., Liu, Z., Guo, X., & Wen, W. (2020) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1109/dac18072.2020.9218633 · Cited by 5
The Impact of Activation Sparsity on Overfitting in Convolutional Neural Networks
Huesmann, K., Rodríguez, L. G., Linsen, L., & Risse, B. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-68796-0_10 · Cited by 10
Li, W., Xu, Z., Sun, Y., Gong, Q., Chen, Y., Ding, A. Y., Wang, X., & Hui, P. (2021) · IEEE Transactions on Knowledge and Data Engineering · doi:10.1109/tkde.2021.3091503 · Cited by 8
“Deep neural networks are now successful in many fields, but their palatable performance in generalizing may still require some theoretical explanation [49].”
BackgroundSeek Common Ground While Reserving Differences: A Model-Agnostic Module for Noisy Domain Adaptation
Zuo, Y., Yao, H., Zhuang, L., & Xu, C. (2021) · IEEE Transactions on Multimedia · doi:10.1109/tmm.2021.3097495 · Cited by 5
Non Parametric Data Augmentations Improve Deep-Learning based Brain Tumor Segmentation
Ben-Atya, H., Rajchert, O., Goshen, L., & Freiman, M. (2021) · IEEE International Conference on Microwaves, Communications, Antennas and Electronic Systems (COMCAS) · doi:10.1109/comcas52219.2021.9629083 · Cited by 6
Hold-out estimates of prediction models for Markov processes
Garnier, R., Langhendries, R., & Rynkiewicz, J. (2023) · Statistics · doi:10.1080/02331888.2023.2183203 · Cited by 5
Yin, Y., Ayoub, M., Abel, A. K., & Zhang, H. (2022) · Lecture notes in networks and systems · doi:10.1007/978-3-031-16078-3_8 · Cited by 5
A Fast Adaptive Online Gradient Descent Algorithm in Over-Parameterized Neural Networks
Yang, A., Li, D., & Li, G. (2022) · Neural Processing Letters · doi:10.1007/s11063-022-11057-4 · Cited by 5
“Studies demonstrate that the training loss of local models can be (nearly) zero in over-parameterized neural networks [15, 16].”
BackgroundRobust Learning Against Label Noise Based on Activation Trend Tracking
Wang, Y., Zhang, Y., Jiang, Z., Li, Z., Chen, J., & Lu, J. (2022) · IEEE Transactions on Instrumentation and Measurement · doi:10.1109/tim.2022.3219493 · Cited by 6
“Recent studies [7], [8] show that DNNs can easily overfit on noisy labels and this can result in poor generalization performance.”
BackgroundMensah, P. K., & Ayidzoe, M. A. (2022) · International Journal of Computing and Digital Systems · doi:10.12785/ijcds/120104 · Cited by 5
“Overfitting can be monitored theoretically by ensuring that a k-layer deep learning model has knd parameters needed to fit a d-dimensional dataset with n samples [48] perfectly.”
BackgroundEnd-to-End Deep Learning for Stress Recognition Using Remote Photoplethysmography
Zhou, K., Schinle, M., Weimar, S. N., Gerdes, M., Stock, S., & Stork, W. (2022) · IEEE International Conference on Bioinformatics and Biomedicine (BIBM) · doi:10.1109/bibm55620.2022.9995577 · Cited by 5
“Various deep learning network architectures were discussed for the end-to-end classification: five architectures for time-series classification such as Fully Convolutional Network (FCN) [21], Residual Network (RESNET) [22], Multi-Layer Perception (MLP) [21], Time Convolutional Neural Network (CNN), and Multichannel Deep…”
MethodCausal Unstructured Pruning in Linear Networks Using Effective Information
Zeng, C., Liu, L., Zhao, H., Zhang, Y., Wang, W., Cai, N., & Yue, Y. (2022) · International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery (CyberC) · doi:10.1109/cyberc55534.2022.00056 · Cited by 9
Meta semi-supervised medical image segmentation with label hierarchy
Xu, H. Y., Xie, H., Tan, Q., & Zhang, Y. (2023) · Health Information Science and Systems · doi:10.1007/s13755-023-00222-1 · Cited by 5
“With only a few labeled samples in SSL, DNN models tend to deteriorate into memorizing all the labeled samples [56].”
BackgroundTakamido, R., & Ota, J. (2023) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2023.3296930 · Cited by 5
“The generalizability of demonstration data is a well-known problem; however, if noisy and small-sized data are directly learned and generalized, the generalization error increases [11].”
BackgroundSharpness-Aware Minimization with Dynamic Reweighting
Zhou, W., Liu, F., Zhang, H., & Chen, M. (2022) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2022.findings-emnlp.417 · Cited by 8
“, 2012), they are often overparameterized and can easily overfit the training data (Zhang et al., 2021).”
BackgroundAutoMRM: A Model Retrieval Method Based on Multimodal Query and Meta-learning
Li, Z., Qi, B., Sun, H., & Gao, X. (2023) · ACM International Conference on Information and Knowledge Management (CIKM) · doi:10.1145/3583780.3614787 · Cited by 5
“For 45 binary image datasets from CIFAR-10, we use six candidate DNN models,the same as in previous work [5]: AlexNet1 [20], MiniGoogLeNet1 [48], MiniVGG1, MiniVGG3 [38], ResNet10 [9] and SmallSqueezeNet [14].”
MethodCMC: Few-shot Novel View Synthesis via Cross-view Multiplane Consistency
Zhu, H., & Chen, Z. (2024) · IEEE Conference on Virtual Reality and 3D User Interfaces (VR) · doi:10.1109/vr58804.2024.00115 · Cited by 7
“To tackle this challenge, we make an assumption: due to fewer input views, the sampling point in each ray would rarely be used to render other views, therefore the neural networks tend to memorize colors of each input view instead of learning the underlying geometry [2,58].”
BackgroundSize and Depth of Monotone Neural Networks: Interpolation and Approximation
Mikulincer, D., & Reichman, D. (2024) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2024.3387878 · Cited by 7
“There is extensive research regarding the size and depth needed for general datasets and networks to achieve interpolation [7], [13], [45], [49] starting with the seminal work of Baum [4].”
Envisioning future deep learning theories: some basic concepts and characteristics
Su, W. (2024) · Science China Information Sciences · doi:10.1007/s11432-023-4129-1 · Cited by 6
“Conventional wisdom from statistical learning theory suggests that a model may not perform well on test data if its parameters outnumber the training samples; to avoid overfit-ting, explicit regularization is needed to constrain the search space of the unknown parameters [30].”
BackgroundData-Aware Customization of Activation Functions Reduces Neural Network Error
Gao, F., & Zhang, B. (2025) · Communications in computer and information science · doi:10.1007/978-3-031-99586-6_9 · Cited by 6
Towards Data-Algorithm Dependent Generalization: a Case Study on Overparameterized Linear Regression
Xu, J., Teng, J., Yuan, Y., & Yao, A. C.-C. (2023) · neural information processing systems · doi:10.52202/075280-3489 · Cited by 12
“One of the most popular data-dependent methods is uniform convergence [7, 29, 74, 75].”
BackgroundWide stochastic networks: Gaussian limit and PAC-Bayesian trainingPreprint
Clerico, E., Deligiannidis, G., & Arnaud Doucet, B. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2106.09798 · Cited by 5
“The complexity bounds from the classical statistical learning theory are manifestly inadequate tools to study the generalization properties of wide and deep neural networks [1].”
BackgroundThe Separation Capacity of Random Neural NetworksPreprint
Dirksen, S., Genzel, M., Jacques, L., & Stollenwerk, A. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2108.00207 · Cited by 5
“The ability of memorizing large data sets (including their noisy components) is a well-known phenomenon in deep learning research and considered as an important piece of the still unsolved generalization puzzle (Zhang et al., 2017, 2021).”
BackgroundCertifying Robustness to Programmable Data Bias in Decision TreesPreprint
Meyer, A. P., Albarghouthi, A., & D’Antoni, L. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2110.04363 · Cited by 26
“Most works have focused on attacks [6, 11, 25, 33, 38, 39, 40] or on training models that are empirically less vulnerable (defenses) [3, 9, 19, 28, 31, 36].”
BackgroundZero to One: Sales Prospecting with Augmented RecommendationPreprint
Hu, S., Zhang, J., & Zhu, Y. (2022) · SSRN Electronic Journal · doi:10.2139/ssrn.4006841 · Cited by 6
“In addition, overparameterized deep learning models may overfit the training data and fail to generalize out-of-sample (Zhang et al. 2021).”
BackgroundMethodInfluentialTorrisi, M., de León, A. D. L. V., Climent, G., Loos, R., & Panjkovich, A. (2022) · bioRxiv (Cold Spring Harbor Laboratory) · doi:10.1101/2022.04.20.488898 · Cited by 5
Why Deep Learning GeneralizesPreprint
Badger, B. L. (2022) · arXiv (Cornell University) · doi:10.48550/arxiv.2211.09639 · Cited by 6
“These observations have led to the hypothesis that it is not model capacity as it is traditionally understood but some other regularizer that prevents deep learning models from overfitting even when they are capable of doing so (Zhang et al., 2021).”
BackgroundHow does training shape the Riemannian geometry of neural network representations?Preprint
Zavatone-Veth, J. A., Yang, S., Rubinfien, J. A., & Pehlevan, C. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2301.11375 · Cited by 8
Privacy Threats in Stable Diffusion ModelsPreprint
Cilloni, T., Fleming, C., & Walter, C. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2311.09355 · Cited by 5
“Zhang et al. [35] observed that standard neural networks can memorize datasets with randomly assigned labels.”
BackgroundAI-Generated Videos and Deepfakes: A Technical PrimerPreprint
Sufian, A. (2024) · doi:10.36227/techrxiv.172348990.01007128/v1 · Cited by 6
ClassiFIM: An Unsupervised Method To Detect Phase TransitionsPreprint
Kasatkin, V., Mozgunov, E., Ezzell, N., Mishra, U., Hen, I., & Lidar, D. A. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2408.03323 · Cited by 5
“…overparameterized models to achieve near-zero cross-entropy loss on the training set (which is possible if all samples x are distinct; see, e.g., Zhang et al. (2021); Ishida et al. (2020)), which leads to inflated absolute values of the log odds and biased estimates ˆ g (i.e., typically, ˆ g >…”
BackgroundTraining on Irrelevant States Implies Data Augmentation: Generalization in Contextual MDPsPreprint
Weltevrede, M., Horsch, C., Spaan, M. T. J., & Böhmer, W. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2410.03565 · Cited by 5
“Data augmentation has proven to be a powerful tool for addressing such spurious correlations and improving generalization performance across a wide range of settings and applications (Shorten & Khoshgoftaar, 2019; Feng et al., 2021; Zhang et al., 2021a).”
Replay Can Provably Increase ForgettingPreprint
Mahdaviyeh, Y., Lucas, J., Ren, M., Tolias, A. S., Zemel, R. S., & Pitassi, T. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2506.04377 · Cited by 6
“It is known that training with GD or SGD leads to a solution that has minimum distance to initialization (Gunasekar et al., 2018; Zhang et al., 2021), that is, (1) The parameter error of the procedure in Equation 1 satisfies the following recursive relationship We include a derivation of this…”
Alaryngeal Speech Enhancement for Noisy Environments Using a Pareto Denoising Gated LSTM
Maskeliūnas, R., Damaševičius, R., Kulikajevas, A., Pribuišis, K., & Uloza, V. (2024) · Journal of Voice · doi:10.1016/j.jvoice.2024.07.016 · Cited by 4
Reimagining cancer treatments in the era of generative AI
Derbal, Y. (2025) · DIGITAL HEALTH · doi:10.1177/20552076251394631 · Cited by 4
“Despite these challenges and the lack of a decisive scienti fi c understanding of why GenAI works so well, 67,68 there is a fl urry of studies exploring the utility of GenAI to support the tasks of clinical oncology, including treatment decision-making.”
Predicting the stress-strain behavior of gravels with a hybrid deep learning approach
Li, D., Liu, J., Zou, D., Xu, K., Ning, F., & Cui, G. (2025) · Transportation Geotechnics · doi:10.1016/j.trgeo.2025.101492 · Cited by 4
Gao, J., Tang, H., Wang, Z., Li, Y., Luo, N., Song, M. X., Xie, S., Shi, W., Yan, H., Lü, L., et al. (2025) · Neuroscience Bulletin · doi:10.1007/s12264-025-01385-5 · Cited by 8
“However, the inherent non-linearity of the GAT model poses challenges in comprehending its internal dynamics and decision-making processes, thereby rendering the model a “black box” that is difficult to explain and comprehend [42].”
Anwar, S., Yunker, A., Kettimuthu, R., Anastasio, M. A., Liu, Z., Su, W., & He, J. (2025) · Smart Materials and Structures · doi:10.1088/1361-665x/adc359 · Cited by 6
“Establishing the model’s generalizability presents notable complexities, particularly in the context of supervised learning, as exemplified in this study [62].”
Yuan, X., Guo, X., Luo, Y., Guan, X., Li, Q., Situ, Z., Zhou, Z., Huang, X., Rong, Z., Lin, Y., et al. (2025) · IEEE Transactions on Medical Imaging · doi:10.1109/tmi.2025.3555621 · Cited by 6
“Studies have shown that deep learning-based models, which is a nonlinear dimensionality reduction, have a stronger generalization ability than traditional ML models in multi-center environments [55], [56].”
Data Glitches Discovery using Influence-based Model Explanations
Myrtakis, N., Tsamardinos, I., & Christophides, V. (2025) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3690624.3709285 · Cited by 5
“This is particularly important for deep neural networks that can overfit a training set even with a high ratio of corrupted labels due to a large number of parameters, resulting in poor generalization performance [57].”
Meng, L., Zhu, W., Xu, N., & Hower, J. C. (2025) · Fuel · doi:10.1016/j.fuel.2025.136072 · Cited by 4
Unequal spatio-temporal distribution of population-weighted pollution extremes through deep learning
Rautela, K. S., Goyal, M. K., & Nagpure, A. S. (2025) · npj Climate and Atmospheric Science · doi:10.1038/s41612-025-01183-w · Cited by 4
“Despite themodel ’ s strong performance, a slight decline in accuracy is observed in the testing phase, suggesting that the model might struggle with generalisation when faced with unseen data 45 .”
Dai, Z., Chen, J., Liu, C., Gao, X., & Yang, L. (2025) · Frontiers in Marine Science · doi:10.3389/fmars.2025.1687230 · Cited by 4
Cheng, Y., Wei, L., & Ding, C. (2025) · Computer Methods in Applied Mechanics and Engineering · doi:10.1016/j.cma.2025.118511 · Cited by 4
Noise-induced equalization in quantum learning models
Scala, F., Guarnieri, G., & Lucchi, A. (2026) · Quantum Science and Technology · doi:10.1088/2058-9565/ae9265 · Cited by 6
“In both classical and quantum settings, bounding the generalization gap via Rademacher complexity or related tools (such as VC dimension or covering numbers [37]) is common practice for designing architectures that generalize well, even if this bounds have been shown to be somewhat vacuous [85].”
Kulkarni, M., Joshi, D., & Patel, S. (2025) · Journal of Pharmaceutical Innovation · doi:10.1007/s12247-025-10236-9 · Cited by 4
Zhou, L., Cheng, X., Liu, S., He, C., Peng, W., & Zhang, M. (2025) · Forests · doi:10.3390/f16121778 · Cited by 4
“When model complexity increases and training epochs accumulate, the continuous gradual decline in training loss may paradoxically weaken generalization performance [73,74].”
Tang, Y. A., Ni, L., Zhang, J., & Liang, Y. (2026) · IEEE Internet of Things Journal · doi:10.1109/jiot.2026.3653444 · Cited by 6
SegDINO: Introducing Multi-scale Structure Into DINO for Efficient Medical Image Segmentation
Yang, S., Wang, H., Xing, Z., Chen, S., Yang, Q., Mao, Y., Yang, G., & Zhu, L. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-38085-2_48 · Cited by 4
Mubarok, R. (2026) · J-TELITE: Journal of Transforming Education through Leadership, Innovation, and Teaching Excellence · doi:10.58586/zek9ws38 · Cited by 4
Convolutional neural network pruning based on misclassification cost
Ahmadluei, S., Faez, K., & Masoumi, B. (2023) · The Journal of Supercomputing · doi:10.1007/s11227-023-05487-7 · Cited by 5
“Surprisingly, acceptable generalization is observed in these networks, and there is no overarching theory that can explain this experimental craft [9–11].”
BackgroundCoSaR: Combating Label Noise Using Collaborative Sample Selection and Adversarial Regularization
Zhang, X., Liu, Y., Wang, H., Wang, W., Ni, P., & Zhang, J. (2023) · ACM International Conference on Information and Knowledge Management (CIKM) · doi:10.1145/3583780.3614826 · Cited by 4
A framework for generalizable neural networks for robust estimation of eyelids and pupils
Biswas, A., & Lescroart, M. D. (2023) · Behavior Research Methods · doi:10.3758/s13428-023-02266-3 · Cited by 5
Optimizer’s dilemma: optimization strongly influences model selection in transcriptomic prediction
Crawford, J., Chikina, M. D., & Greene, C. S. (2024) · Bioinformatics Advances · doi:10.1093/bioadv/vbae004 · Cited by 4
Does Confusion Really Hurt Novel Class Discovery?
Chi, H., Yang, W., Liu, T., Lan, L., Qin, T., & Han, B. (2024) · International Journal of Computer Vision · doi:10.1007/s11263-024-02012-y · Cited by 4
“Due to strong memory ability of deep networks (Zhang et al., 2021), the learning errors caused by noisy supervision signals will persistently accumulate during training.”
SLaNT: A Semi-supervised Label Noise-Tolerant Framework for Text Sentiment Analysis
Cao, B., Jiang, K., & Fan, J. (2024) · Proceedings of the International AAAI Conference on Web and Social Media · doi:10.1609/icwsm.v18i1.31307 · Cited by 5
“However, due to the memorization effect (Zhang et al. 2021a), DNN model will overfit noisy data, subsequently impacting the performance of CL.”
BackgroundMethodA Noisy Sample Selection Framework Based on a Mixup Loss and Recalibration Strategy
Zhang, Q., Yu, D., Zhou, X., Gong, H., Li, Z., Liu, Y., & Shao, R. (2024) · Mathematics · doi:10.3390/math12152389 · Cited by 4
Berthelot, L., Zavagno, A., Artières, T., Dupé, F.-X., Gray, M., Russeil, D., Schisano, E., & Arzoumanian, D. (2024) · Astronomy and Astrophysics · doi:10.1051/0004-6361/202450828 · Cited by 6
Recent advances in complementary label learning
Tian, Y., & Jiang, H. (2024) · Information Fusion · doi:10.1016/j.inffus.2024.102702 · Cited by 4
Can the accuracy bias by facial hairstyle be reduced through balancing the training data?
Öztürk, K., Wu, H., & Bowyer, K. W. (2024) · IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · doi:10.1109/cvprw63382.2024.00159 · Cited by 4
“Bias towards low complexity solutions in deep networks are investigated in recent works [5,27,28,35,46,50].”
Impact of Class Imbalance on Unsupervised Label Generation for Medicare Fraud Detection
Kennedy, K. L. R., & Khoshgoftaar, T. M. (2024) · IEEE International Conference on Information Reuse and Integration (IRI) · doi:10.1109/iri62200.2024.00053 · Cited by 4
On the Generalization and Causal Explanation in Self-Supervised Learning
Qiang, W., Song, Z., Gu, Z., Li, J., Zheng, C., Sun, F., & Xiong, H. (2024) · International Journal of Computer Vision · doi:10.1007/s11263-024-02263-9 · Cited by 5
“Overfitting means that the network memorizes the specific characteristics of the training data instead of learning a generalizable solution (Zhang et al., 2021; Stephenson et al., 2021).”
Impact of Batch Normalization on Convolutional Network Representations
Potgieter, H. L., Mouton, C., & Davel, M. H. (2024) · Communications in computer and information science · doi:10.1007/978-3-031-78255-8_14 · Cited by 4
Learning Non-uniform Step Sizes for Neural Network Quantization
Gongyo, S., Liang, J., Ambai, M., Kawakami, R., & Sato, I. (2024) · Lecture notes in computer science · doi:10.1007/978-981-96-0966-6_4 · Cited by 7
“In these applications, large over-parameterized models can generally achieve high performance [60], but they require huge computational costs during inference.”
Estimated Tardiness-Based Reinforcement Learning Solution to Repeatable Job-Shop Scheduling Problems
Heo, C. Y., Seo, J., Kim, Y., Kim, Y., & Kim, T. (2024) · Processes · doi:10.3390/pr13010062 · Cited by 7
On the effect of noise on fitting linear regression models
Ullah, I., & Welsh, A. H. (2026) · Computational Statistics & Data Analysis · doi:10.1016/j.csda.2026.108421 · Cited by 6
“…with increasing model complexity (Belkin et al. 2019, Geiger et al. 2019, Belkin et al. 2020, Bartlett et al. 2020, Kobak et al. 2020, Liang & Rakhlin 2020, Liang et al. 2020, Muthukumar et al. 2020, Nakkiran et al. 2021, Zhang et al. 2021, Hastie et al. 2022, Deng et al. 2022, Rocks & Mehta 2022).”
BackgroundApproximating Human-Level 3D Visual Inferences With Deep Neural Networks
O’Connell, T. P., Bonnen, T., Friedman, Y., Tewari, A., Sitzmann, V., Tenenbaum, J. B., & Kanwisher, N. (2025) · Open Mind · doi:10.1162/opmi_a_00189 · Cited by 7
Balgi, S., Daoud, A., Peña, J. M., Wodtke, G. T., & Zhou, J. (2025) · Sociological Methods & Research · doi:10.1177/00491241251319291 · Cited by 7
Ecological Toxicity, Oxidative Stress and Impacts of Microplastics on Fish Gills
Waheed, N., Naseer, A., Shabbir, F., Abbas, M., Iqbal, S., Ahmad, S., Bashir, M. S., Mahmood, A., Rizvi, S. M. J. R., Bano, Z., et al. (2025) · Indus journal of bioscience research. · doi:10.70749/ijbr.v3i3.785 · Cited by 4
Knowledge-enhanced ontology-to-vector for automated ontology concept enrichment in BIM
Wei, Y., & Li, X. (2025) · Journal of Industrial Information Integration · doi:10.1016/j.jii.2025.100836 · Cited by 4
Fischer-Brandies, L., Müller, L., Rebholz, B., & Buettner, R. (2025) · IEEE Access · doi:10.1109/access.2025.3556582 · Cited by 5
“This means that while these models demonstrate a high level of performance on the specific data they are trained on (generalization) [37], they perform significantly worse in real-world situations as the real-world input data tends to vary from the training data [38].”
Grey-box model of N2O and NH4+ for predictive control in the activated sludge process
Hansen, L. D., Stentoft, P. A., Stokholm-Bjerregaard, M., Ortíz-Arroyo, D., & Durdevic, P. (2025) · Journal of Water Process Engineering · doi:10.1016/j.jwpe.2025.107503 · Cited by 4
Alagoz, B. B., Keleş, C., Ateş, A., Özdemir, E., & Alkhulaifi, N. (2025) · Neural computing & applications (Print) · doi:10.1007/s00521-025-11212-7 · Cited by 4
“…cutout techniques [24], DropConnect [25] and adaptive DropConnect [26] methods, neural tangent kernels [27], neural network stiffness measures [28], Maxup [15], bias-variance decomposition [16], regularization techniques [29], architecture optimization [30], and neuroevo-lutionary approaches [17].”
AdvMixUp: Adversarial MixUp Regularization for Deep Learning
Fu, J., Ji, X., Chen, D., Hu, G., Li, S., & Feng, X. (2025) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2025.3562363 · Cited by 4
“However, due to ample parameter space and limited training data, DNNs can easily memorize training samples [10], leading to overfitting.”
AI-based geological subsurface reconstruction using sparse convolutional autoencoders
Uribe‐Ventura, R., Barriga-Berrios, Y., Gamarra, J. P. B., Baby, P., & Viveen, W. (2025) · Computers & Geosciences · doi:10.1016/j.cageo.2025.105981 · Cited by 4
Deep Reinforcement Learning Driven Weakly Supervised Lesion Segmentation in Lung CT Sequence Images
Lai, Y., Liu, X., Liang, X., Rong, C., Wu, Y., & Zheng, W. (2025) · IEEE Transactions on Instrumentation and Measurement · doi:10.1109/tim.2025.3571146 · Cited by 5
“To mitigate these challenges, researchers have explored various strategies such as data augmentation [11], [12], [13], transfer learning [14], [15], and domain adaptation [16], [17] to reduce reliance on extensive labeled data.”
Lee, G., Shin, J. Y., & Kim, H. (2025) · Neural Networks · doi:10.1016/j.neunet.2025.107697 · Cited by 6
Song, Z., He, W., Yao, Y., Yu, L., Huang, J., Xu, Y., & Wang, H. (2025) · Frontiers in Plant Science · doi:10.3389/fpls.2025.1620971 · Cited by 4
Metadata
- Authors
- Zhang, C., Bengio, S., Hardt, M., Recht, B., & Vinyals, O.
- Affiliations
- Google (United States) (US)
- Google Brain, Mountain View, CA
- University of California, Berkeley (US)
- Google DeepMind (United Kingdom) (GB)
- Venue
- Communications of the ACM
- Published
- 2021-02-22
- Volume
- 64
- Issue
- 3
- Pages
- 107-115
- ISSN
- 0001-0782 (ISSN-L)
- 1557-7317
- OpenAlex
- W2566079294
- Cited by
- 2,400
- References
- 47
- Citations by year
- 2026: 295 · 2025: 558 · 2024: 579 · 2023: 500 · 2022: 316 · 2021: 108
- FWCI
- 132.37
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- Locations
- Communications of the ACM (publishedVersion, cc-by-nc-sa)
- Funders
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Stochastic Gradient Optimization Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Neural Networks and Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Keywords
- generalization, deep neural network, overfitting, overparameterization, expressivity, stochastic gradient descent, regularization, image classification, convolutional neural networks
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-11-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-11-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.