On the Effects of Artificial Data Modification
Not on the map. Why?
Abstract (arXiv)
Data distortion is commonly applied in vision models during both training (e.g methods like MixUp and CutMix) and evaluation (e.g. shape-texture bias and robustness). This data modification can introduce artificial information. It is often assumed that the resulting artefacts are detrimental to training, whilst being negligible when analysing models. We investigate these assumptions and conclude that in some cases they are unfounded and lead to incorrect results. Specifically, we show current shape bias identification methods and occlusion robustness measures are biased and propose a fairer alternative for the latter. Subsequently, through a series of experiments we seek to correct and strengthen the community's perception of how augmenting affects learning of vision models. Based on our empirical results we argue that the impact of the artefacts must be understood and exploited rather than eliminated.
Key facts. Cited by 2 papers in the map (data through 14 Oct 2026). Builds on 33 of its 44 references that are in the map. First posted on arXiv 26 Oct 2021.
Builds on
33 of 44 references.
References · 33
- 2014Very Deep Convolutional Networks for Large-Scale Image Recognitioncited by 15,364
- 2020A Simple Framework for Contrastive Learning of Visual Representationscited by 12,049
- 2014ImageNet Large Scale Visual Recognition Challengecited by 8,498
- 2019Momentum Contrast for Unsupervised Visual Representation Learningcited by 7,513
- 2014Explaining and Harnessing Adversarial Examplescited by 7,364
- 2016Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localizationcited by 6,733
- 2016Densely Connected Convolutional Networkscited by 6,720
- 2017mixup: Beyond Empirical Risk Minimizationcited by 5,183
- 2013Intriguing properties of neural networkscited by 5,155
- 2017Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithmscited by 4,349
- 2016Wide Residual Networkscited by 3,070
- 2019CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Featurescited by 2,884
21 more
- 2016Identity Mappings in Deep Residual Networkscited by 1,965
- 2017Improved Regularization of Convolutional Neural Networks with Cutoutcited by 1,628
- 2018ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustnesscited by 1,432
- 2017Random Erasing Data Augmentationcited by 1,239
- 2020Captum: A unified and generic model interpretability library for PyTorchcited by 547
- 2020What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimationcited by 497
- 2018A Benchmark for Interpretability Methods in Deep Neural Networkscited by 476
- 2019Domain Generalization by Solving Jigsaw Puzzlescited by 450
- 2019Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNetcited by 260
- 2020On Mixup Regularizationcited by 83
- 2019Interpreting Adversarially Trained Convolutional Neural Networkscited by 75
- 2020Informative Dropout for Robust Representation Learning: A Shape-bias Perspectivecited by 57
- 2020Affinity and Diversity: Quantifying Mechanisms of Data Augmentationcited by 45
- 2018Improved Mixed-Example Data Augmentationcited by 42
- 2021Does enhanced shape bias improve neural network robustness to common corruptions?cited by 32
- 2020Representation Based Complexity Measures for Predicting Generalization in Deep Learningcited by 29
- 2020An Empirical Evaluation on Robustness and Uncertainty of Regularization Methodscited by 28
- 2017Recurrent computations for visual pattern completioncited by 22
- 2020FMix: Enhancing Mixed Sample Data Augmentationcited by 11
- 2020Ranking Deep Learning Generalization using Label Variation in Latent Geometry Graphscited by 6
- 2021Gi and Pal Scores: Deep Neural Network Generalization Statisticscited by 3
Built on by
Citing papers per year: 2022: 1; 2023: 1.
2 papers.