pith. sign in

arxiv: 1708.04552 · v2 · submitted 2017-08-15 · 💻 cs.CV

Improved Regularization of Convolutional Neural Networks with Cutout

Terrance DeVries , Graham W. Taylor This is my paper

Pith reviewed 2026-05-13 20:32 UTC · model grok-4.3

classification 💻 cs.CV
keywords cutoutregularizationconvolutional neural networksdata augmentationCIFAR-10CIFAR-100SVHNoverfitting
0
0 comments X

The pith

Randomly masking square regions in training images improves convolutional neural network generalization.

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

The paper introduces cutout, a regularization method that randomly masks out square patches from input images during training. This forces convolutional networks to learn more distributed and robust features instead of relying on specific local patterns. The approach requires no architectural changes and combines readily with standard data augmentation. Experiments on CIFAR-10, CIFAR-100, and SVHN show it produces new state-of-the-art error rates when added to existing high-performing models. Readers would care because it offers a low-effort way to reduce overfitting in image classification tasks.

Core claim

The paper claims that randomly masking fixed-size square regions of the input image during training, called cutout, acts as an effective regularizer that improves the robustness and test accuracy of convolutional neural networks, achieving 2.56% error on CIFAR-10, 15.20% on CIFAR-100, and 1.30% on SVHN when applied to current state-of-the-art architectures.

What carries the argument

Cutout: the operation of selecting a random square region and setting its pixels to zero in each training image to encourage feature robustness.

Load-bearing premise

A single fixed square size and random placement will produce consistent gains across architectures and datasets without requiring dataset-specific retuning or introducing harmful bias in the learned features.

What would settle it

Applying cutout with one fixed mask size to ImageNet or another large-scale dataset and observing no reduction in top-1 error relative to the unaugmented baseline would show the gains do not generalize.

read the original abstract

Convolutional neural networks are capable of learning powerful representational spaces, which are necessary for tackling complex learning tasks. However, due to the model capacity required to capture such representations, they are often susceptible to overfitting and therefore require proper regularization in order to generalize well. In this paper, we show that the simple regularization technique of randomly masking out square regions of input during training, which we call cutout, can be used to improve the robustness and overall performance of convolutional neural networks. Not only is this method extremely easy to implement, but we also demonstrate that it can be used in conjunction with existing forms of data augmentation and other regularizers to further improve model performance. We evaluate this method by applying it to current state-of-the-art architectures on the CIFAR-10, CIFAR-100, and SVHN datasets, yielding new state-of-the-art results of 2.56%, 15.20%, and 1.30% test error respectively. Code is available at https://github.com/uoguelph-mlrg/Cutout

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper introduces Cutout, a simple regularization technique for CNNs that randomly masks out square regions of the input image during training. The authors show that Cutout can be combined with existing data augmentations and regularizers, and report new state-of-the-art test errors of 2.56% on CIFAR-10, 15.20% on CIFAR-100, and 1.30% on SVHN when applied to modern architectures. Public code is released for reproducibility.

Significance. If the results hold, the work is significant because it supplies an extremely lightweight regularization method that delivers consistent gains on top of strong baselines and yields new SOTA numbers on three standard benchmarks. The code release is a clear strength, supporting verification and further use by the community.

major comments (1)
  1. [Experiments] Experiments section: the reported SOTA results use manually selected fixed cutout sizes (16×16 on CIFAR-10/100, 20×20 on SVHN). No systematic sensitivity sweep or cross-architecture transfer experiment is presented to show that performance gains hold for a range of sizes without per-dataset retuning. This leaves open whether the improvements are attributable to the method itself or to implicit hyper-parameter selection.
minor comments (2)
  1. [Abstract] The abstract states that Cutout is applied to 'current state-of-the-art architectures' but does not name the specific models used for each dataset; adding this detail would improve clarity.
  2. [Method] In the method description, the precise implementation of the mask (e.g., whether it is applied identically across all channels and how boundary handling is performed) could be stated more explicitly to facilitate exact reproduction.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive review and the recommendation for minor revision. We appreciate the recognition of Cutout as a lightweight regularization technique that yields consistent gains and new state-of-the-art results. We address the single major comment below.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: the reported SOTA results use manually selected fixed cutout sizes (16×16 on CIFAR-10/100, 20×20 on SVHN). No systematic sensitivity sweep or cross-architecture transfer experiment is presented to show that performance gains hold for a range of sizes without per-dataset retuning. This leaves open whether the improvements are attributable to the method itself or to implicit hyper-parameter selection.

    Authors: We thank the referee for highlighting this point. The cutout sizes were selected as roughly half the side length of the input images (CIFAR and SVHN images are 32×32 pixels), which provides a natural scale for occluding a meaningful portion of the image without removing all semantic content. While the original manuscript focused on demonstrating the method's effectiveness when combined with strong modern architectures and existing augmentations, we agree that an explicit sensitivity analysis would better substantiate that the gains arise from the regularization mechanism itself rather than from dataset-specific tuning. In the revised version we will add a new figure and accompanying text in the Experiments section that reports test error on CIFAR-10 (using Wide ResNet-28-10) for cutout sizes ranging from 0 to 24 pixels in steps of 4. This will show that performance improvements are obtained across a broad interval around the chosen size of 16, thereby addressing the concern about implicit hyper-parameter selection. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical results rest on independent training runs with no derivations or self-referential reductions.

full rationale

This is an empirical paper proposing the Cutout regularization method (random square masking of inputs) and reporting its effect when combined with existing augmentations. All performance numbers (2.56% on CIFAR-10, 15.20% on CIFAR-100, 1.30% on SVHN) are obtained from explicit model training and evaluation on held-out test sets; no equations, fitted parameters, or predictive derivations appear in the work. Consequently there are no self-definitional steps, fitted-input-called-prediction steps, or load-bearing self-citations that collapse any claim to its own inputs by construction. The method's description and experimental protocol are self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The claim rests on the empirical effectiveness of random square masking as a regularizer. The only notable free parameter is the cutout square size, which is chosen per dataset. No new entities are postulated and no circular derivations appear.

free parameters (1)
  • cutout square size
    The side length of the masked square is a hyperparameter tuned separately for each dataset and architecture.
axioms (1)
  • domain assumption Standard CNN training assumptions (SGD, cross-entropy loss, data augmentation pipeline)
    The method is applied on top of existing training procedures without altering their core assumptions.

pith-pipeline@v0.9.0 · 5477 in / 1330 out tokens · 47010 ms · 2026-05-13T20:32:05.432740+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 41 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Navigating Potholes with Geometry-Aware Sharpness Minimization

    cs.LG 2026-05 unverdicted novelty 7.0

    LLQR+SAM pairs a slow learned geometry preconditioner with fast SAM perturbations to amplify escape from locally sharp 'potholes' while stabilizing flat basins, producing consistent gains over SAM and LLQR alone.

  2. Embracing Biased Transition Matrices for Complementary-Label Learning with Many Classes

    cs.LG 2026-05 unverdicted novelty 7.0

    BICL uses biased non-uniform transition matrices to generate constrained complementary labels, enabling effective learning and over sevenfold accuracy gains on many-class image datasets.

  3. Characterizing the Generalization Error of Random Feature Regression with Arbitrary Data-Augmentation

    stat.ML 2026-05 conditional novelty 7.0

    The test error of random-feature ridge regression with arbitrary data augmentation admits a closed-form asymptotic characterization in the proportional regime that depends only on population covariances and augmentati...

  4. SeBA: Semi-supervised few-shot learning via Separated-at-Birth Alignment for tabular data

    cs.LG 2026-05 unverdicted novelty 7.0

    SeBA is a joint-embedding framework that separates tabular data into two complementary views and aligns one view's representations to the nearest-neighbor structure of the other, improving feature-label relationships ...

  5. Layerwise LQR for Geometry-Aware Optimization of Deep Networks

    cs.LG 2026-05 unverdicted novelty 7.0

    Steepest descent under divergence-induced quadratic models equals an LQR problem, enabling learning of diagonal or Kronecker-factored inverse preconditioners via a global layerwise objective for scalable geometry-awar...

  6. QB-LIF: Learnable-Scale Quantized Burst Neurons for Efficient SNNs

    cs.CV 2026-04 unverdicted novelty 7.0

    QB-LIF uses a trainable quantization scale for burst neurons in SNNs to raise accuracy at ultra-low latency on vision and event datasets while preserving neuromorphic hardware compatibility.

  7. Channel-Level Semantic Perturbations: Unlearnable Examples for Diverse Training Paradigms

    cs.LG 2026-04 unverdicted novelty 7.0

    Unlearnable examples fail under pretraining-finetuning due to semantic filtering by frozen layers, but Shallow Semantic Camouflage restores effectiveness by confining perturbations to semantically valid subspaces.

  8. Seeing Through the Tool: A Controlled Benchmark for Occlusion Robustness in Foundation Segmentation Models

    cs.CV 2026-04 unverdicted novelty 7.0

    SAM-family models split into occluder-aware types that avoid predicting into occluded regions and occluder-agnostic types that confidently segment hidden areas, shown via a new benchmark on polyp datasets.

  9. Perturb and Recover: Fine-tuning for Effective Backdoor Removal from CLIP

    cs.LG 2024-12 conditional novelty 7.0

    PAR fine-tunes CLIP to remove backdoors from structured triggers while preserving standard performance, and works even with only synthetic image-text pairs.

  10. A Simple Framework for Contrastive Learning of Visual Representations

    cs.LG 2020-02 accept novelty 7.0

    SimCLR learns visual representations by contrasting augmented views of the same image and reaches 76.5% ImageNet top-1 accuracy with a linear classifier, matching a supervised ResNet-50.

  11. Point Cloud Sequence Encoding for Material-conditioned Graph Network Simulators

    cs.LG 2026-05 unverdicted novelty 6.0

    PEACH uses a novel spatio-temporal point cloud sequence encoder plus auxiliary supervision to enable zero-shot adaptation of graph network simulators to unseen physical properties, outperforming mesh-based baselines i...

  12. Anatomy of a failure: When, how, and why deep vision fails in scientific domains

    cs.CV 2026-05 unverdicted novelty 6.0

    Deep learning on information-rich scientific images collapses to one-dimensional predictions due to a mismatch between data priors and the model's simplicity bias, even after robustification techniques.

  13. IonMorphNet: Generalizable Learning of Ion Image Morphologies for Peak Picking in Mass Spectrometry Imaging

    cs.CV 2026-04 unverdicted novelty 6.0

    IonMorphNet is a ConvNeXt-based classifier trained on six spatial pattern classes from 53 MSI datasets that performs generalizable peak picking and improves mSCF1 by 7% over prior methods while also aiding tumor class...

  14. Enhancing Tabular Anomaly Detection via Pseudo-Label-Guided Generation

    cs.AI 2026-04 unverdicted novelty 6.0

    PLAG boosts tabular anomaly detection by using pseudo-label-guided synthetic anomaly generation with a two-stage filter, achieving SOTA results and lifting F1 scores by 0.08-0.21 when added to existing detectors.

  15. Soft Label Pruning and Quantization for Large-Scale Dataset Distillation

    cs.CV 2026-04 unverdicted novelty 6.0

    LPQLD reduces soft label storage in dataset distillation by 78-500x on ImageNet datasets via pruning with dynamic reuse and quantization with student-teacher alignment, while improving accuracy.

  16. FireSenseNet: A Dual-Branch CNN with Cross-Attentive Feature Interaction for Next-Day Wildfire Spread Prediction

    cs.CV 2026-04 unverdicted novelty 6.0

    FireSenseNet dual-branch CNN with CAFIM cross-attention outperforms larger models on next-day wildfire spread prediction, reaching F1 of 0.4176 on the Google benchmark.

  17. OASIC: Occlusion-Agnostic and Severity-Informed Classification

    cs.CV 2026-04 conditional novelty 6.0

    OASIC uses anomaly-based masking and severity estimation to select occlusion-matched models, improving AUC on occluded images by up to 23.7 points.

  18. Semantic-aware Random Convolution and Source Matching for Domain Generalization in Medical Image Segmentation

    cs.CV 2025-12 unverdicted novelty 6.0

    Semantic-aware random convolution and intensity-based source matching enable effective single-source domain generalization for medical image segmentation, outperforming prior methods and sometimes matching in-domain p...

  19. Masked Language Prompting for Generative Data Augmentation in Few-shot Fashion Style Recognition

    cs.CV 2025-04 unverdicted novelty 6.0

    Masked Language Prompting masks selected words in reference captions and leverages LLMs to produce diverse, semantically coherent completions for style-consistent generative image augmentation without fine-tuning.

  20. Orthogonal Subspace Decomposition for Generalizable AI-Generated Image Detection

    cs.CV 2024-11 unverdicted novelty 6.0

    Orthogonal subspace decomposition via SVD on vision foundation model features preserves high-rank pre-trained knowledge by freezing principal components and adapting residuals, reducing overfitting for better generali...

  21. Decouple then Converge: Handling Unknown Unlabeled Distributions in Long-Tailed Semi-Supervised Learning

    cs.LG 2024-06 unverdicted novelty 6.0

    DeCon decouples LTSSL into head-class and tail-class branches that interact and converge, delivering SOTA accuracy on mismatched-distribution benchmarks and outperforming prior methods even on matched distributions.

  22. Sharpness-Aware Minimization for Efficiently Improving Generalization

    cs.LG 2020-10 conditional novelty 6.0

    SAM solves a min-max problem to locate flat low-loss regions, improving generalization on CIFAR, ImageNet and label-noise tasks.

  23. DropAttention: A Regularization Method for Fully-Connected Self-Attention Networks

    cs.CL 2019-07 unverdicted novelty 6.0

    DropAttention regularizes attention weights in fully-connected self-attention networks to reduce overfitting and improve performance.

  24. XferNAS: Transfer Neural Architecture Search

    cs.LG 2019-07 unverdicted novelty 6.0

    XferNAS transfers knowledge across neural architecture searches to reduce search time by a factor of 33 on CIFAR-10/100 while achieving new records of 1.99% and 14.06% error.

  25. Learning Data Augmentation Strategies for Object Detection

    cs.CV 2019-06 unverdicted novelty 6.0

    Learned data augmentation policies optimized for object detection improve COCO mAP by more than 2.3 and transfer to other datasets and models.

  26. Dual-Prompt CLIP with Hybrid Visual Encoders for Occluded Person Re-Identification

    cs.CV 2026-05 unverdicted novelty 5.0

    DPL-ReID adds dual prompt learning, real-world occlusion augmentation, and weighted gated fusion to CLIP for state-of-the-art occluded person re-identification on benchmark datasets.

  27. Margin-Adaptive Confidence Ranking for Reliable LLM Judgement

    cs.LG 2026-05 unverdicted novelty 5.0

    Introduces a margin-adaptive confidence ranking method that learns an estimator from simulated diversity and derives margin-dependent generalization bounds for use in fixed-sequence testing of LLM-human agreement.

  28. ZScribbleSeg: A comprehensive segmentation framework with modeling of efficient annotation and maximization of scribble supervision

    cs.CV 2026-05 unverdicted novelty 5.0

    ZScribbleSeg maximizes scribble supervision with efficient annotation forms, spatial regularization, and EM-estimated class ratios to deliver competitive performance on six medical segmentation tasks without full labels.

  29. Accuracy Improvement of Semi-Supervised Segmentation Using Supervised ClassMix and Sup-Unsup Feature Discriminator

    cs.CV 2026-04 unverdicted novelty 5.0

    Supervised ClassMix and a Sup-Unsup Feature Discriminator yield an average 2.07% mIoU gain over standard semi-supervised methods on Chase and COVID-19 datasets.

  30. Bi-Level Optimization for Single Domain Generalization

    cs.LG 2026-04 unverdicted novelty 5.0

    BiSDG applies bi-level optimization with surrogate domains and a domain prompt encoder to achieve state-of-the-art results in single domain generalization.

  31. WRF4CIR: Weight-Regularized Fine-Tuning Network for Composed Image Retrieval

    cs.CV 2026-04 unverdicted novelty 5.0

    WRF4CIR uses weight-regularized fine-tuning with adversarial perturbations to mitigate overfitting in composed image retrieval and narrows the generalization gap on benchmarks.

  32. Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It

    eess.IV 2026-04 unverdicted novelty 5.0

    MaskGen improves domain generalization for biomedical image segmentation by using source intensities plus domain-stable foundation model representations with minimal added complexity.

  33. YOLOv4: Optimal Speed and Accuracy of Object Detection

    cs.CV 2020-04 unverdicted novelty 5.0

    YOLOv4 achieves 43.5% AP (65.7% AP50) on MS COCO at ~65 FPS on Tesla V100 by integrating WRC, CSP, CmBN, SAT, Mish activation, Mosaic augmentation, DropBlock, and CIoU loss.

  34. How Data Augmentation Shapes Neural Representations

    cs.LG 2026-05 unverdicted novelty 4.0

    Data augmentation produces well-behaved trajectories in shape-invariant representation space, with augmentation type steering distinct directions and geometry predicting ensembling gains.

  35. AtteConDA: Attention-Based Conflict Suppression in Multi-Condition Diffusion Models and Synthetic Data Augmentation

    cs.CV 2026-05 unverdicted novelty 4.0

    AtteConDA adds attention-based conflict suppression to multi-condition diffusion models so that generated driving-scene images retain richer structural cues from the original annotations.

  36. FGML-DG: Feynman-Inspired Cognitive Science Paradigm for Cross-Domain Medical Image Segmentation

    cs.CV 2026-04 unverdicted novelty 4.0

    FGML-DG applies Feynman-inspired principles of concept simplification, memory recall, and error-focused retraining within a meta-learning setup to enhance domain generalization for medical image segmentation.

  37. Single-bit-per-weight deep convolutional neural networks without batch-normalization layers for embedded systems

    cs.LG 2019-07 unverdicted novelty 4.0

    Experiments show that shifted-ReLU layers can replace batch-normalization in single-bit-weight wide residual networks on CIFAR-10/100 and ImageNet without consistent accuracy penalty.

  38. Further advantages of data augmentation on convolutional neural networks

    cs.CV 2019-06 unverdicted novelty 4.0

    Data augmentation enables CNNs to adapt to varying architectures and data amounts without hyperparameter fine-tuning, unlike weight decay and dropout.

  39. SoK: A Comprehensive Analysis of the Current Status of Neural Tangent Generalization Attacks with Research Directions

    cs.LG 2026-05 accept novelty 3.0

    NTGA is the first clean-label generalization attack under black-box settings but is vulnerable to adversarial training and image transformations, with newer attacks outperforming it.

  40. Data-Centric Foundation Models in Computational Healthcare: A Survey

    cs.LG 2024-01 unverdicted novelty 3.0

    The paper surveys data-centric strategies for foundation models in computational healthcare and supplies a curated list of related models and datasets.

  41. Genetic Network Architecture Search

    cs.NE 2019-07 unverdicted novelty 3.0

    Genetic algorithm searches convolution cell architectures with weight sharing via SGD, reporting 96% accuracy on CIFAR10 and 80.1% on CIFAR100.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · cited by 41 Pith papers

  1. [1]

    Bengio, A

    Y . Bengio, A. Bergeron, N. Boulanger-Lewandowski, T. Breuel, Y . Chherawala, et al. Deep learners benefit more from out-of-distribution examples. In Proceedings of the Fourteenth International Conference on Artificial Intelli- gence and Statistics, pages 164–172, 2011

    work page 2011

  2. [2]

    Canziani, A

    A. Canziani, A. Paszke, and E. Culurciello. An analysis of deep neural network models for practical applications. In IEEE International Symposium on Circuits & Systems, 2016

    work page 2016

  3. [3]

    Coates, A

    A. Coates, A. Ng, and H. Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the Fourteenth International Conference on Artificial In- telligence and Statistics, pages 215–223, 2011

    work page 2011

  4. [4]

    Shake-Shake regularization

    X. Gastaldi. Shake-shake regularization. arXiv preprint arXiv:1705.07485, 2017

    work page Pith review arXiv 2017

  5. [5]

    K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in deep residual networks. In European Conference on Com- puter Vision, pages 630–645. Springer, 2016

    work page 2016

  6. [6]

    G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R. Salakhutdinov. Improving neural networks by pre- venting co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580, 2012

    work page Pith review arXiv 2012

  7. [7]

    Krizhevsky and G

    A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. 2009

    work page 2009

  8. [8]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems , pages 1097–1105, 2012

    work page 2012

  9. [9]

    LeCun, L

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner. Gradient- based learning applied to document recognition. Proceed- ings of the IEEE, 86(11):2278–2324, 1998

    work page 1998

  10. [10]

    Lemley, S

    J. Lemley, S. Bazrafkan, and P. Corcoran. Smart augmentation-learning an optimal data augmentation strat- egy. IEEE Access, 2017

    work page 2017

  11. [11]

    J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431– 3440, 2015

    work page 2015

  12. [12]

    Netzer, T

    Y . Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, and A. Y . Ng. Reading digits in natural images with unsupervised fea- ture learning. In NIPS Workshop on Deep Learning and Un- supervised Feature Learning, volume 2011, page 5, 2011

    work page 2011

  13. [13]

    Park and N

    S. Park and N. Kwak. Analysis on the dropout effect in con- volutional neural networks. In Asian Conference on Com- puter Vision, pages 189–204. Springer, 2016

    work page 2016

  14. [14]

    Pathak, P

    D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, and A. A. Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536–2544, 2016

    work page 2016

  15. [15]

    Srivastava, G

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15(1):1929–1958, 2014

    work page 1929

  16. [16]

    Tompson, R

    J. Tompson, R. Goroshin, A. Jain, Y . LeCun, and C. Bregler. Efficient object localization using convolutional networks. In CVPR, pages 648–656, 2015

    work page 2015

  17. [17]

    Toshev and C

    A. Toshev and C. Szegedy. Deeppose: Human pose estima- tion via deep neural networks. In CVPR, pages 1653–1660, 2014

    work page 2014

  18. [18]

    Vincent, H

    P. Vincent, H. Larochelle, I. Lajoie, Y . Bengio, and P.- A. Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local de- noising criterion. Journal of Machine Learning Research , 11(Dec):3371–3408, 2010

    work page 2010

  19. [19]

    Vinyals, A

    O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and tell: A neural image caption generator. In CVPR, pages 3156–3164, 2015

    work page 2015

  20. [20]

    Wu and X

    H. Wu and X. Gu. Towards dropout training for convolu- tional neural networks. Neural Networks, 71:1–10, 2015

    work page 2015

  21. [21]

    R. Wu, S. Yan, Y . Shan, Q. Dang, and G. Sun. Deep image: Scaling up image recognition. arXiv preprint arXiv:1501.02876, 7(8), 2015

    work page Pith review arXiv 2015

  22. [22]

    Zagoruyko and N

    S. Zagoruyko and N. Komodakis. Wide residual networks. British Machine Vision Conference (BMVC), 2016. A. Supplementary Materials 0 20 40 60 80 100 120 Feature/uni00A0activations/uni00A0(sorted/uni00A0by/uni00A0magnitude) 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00Magnitude/uni00A0of/uni00A0activation Cutout Baseline (a) 2nd Residual Block 0 50 100 150 2...

    work page 2016