NeurIPS 2020

The Origins and Prevalence of Texture Bias in Convolutional Neural Networks

Review 1

Summary and Contributions: This paper works to determine the factors that cause current ImageNet-trained CNNs to be biased towards texture. The successfully isolate several factors, and additionally evaluate the bias of non-supervised methods.

Strengths: This is the first principled analysis I know of investigating the phenomenon of texture bias. The writing is very clear, and I found the analysis quite insightful. Overall, it is a very engaging and interesting read. The findings of aggressive random cropping exacerbating texture bias are extremely interesting. I feel that the results of Appendix B could merit inclusion in the main body of the paper. The finding that SimCLR is more biased towards shape, but then recognizing that this is due to the augmentation and then applying said augmentation to supervision is clever.

Weaknesses: The figures generally need much improvement Figure 1: WordNet IDs should be text labels Figure 2: Aesthetically this figure looks very low quality despite the information contained. Error bars shouldn't be the same width as the lines, font is non-standard. Having AlexNet and ResNet50 on different panels makes them hard to compare. I would recommend plotting both architectures on the same subplot Figure 3: The random shape sequence is not very informative. I would recommend color coding the sequence in a rainbow. Font size should be increased for the legend. Figure 4: The font is small enough to be hard to read (overall I do like this finding though) Figure 5: Stylistically many problems are shared with Figure 2. In general Tables besides 1 could have more guiding visuals (e.g. bolding)

Correctness: Yes, I find the experiments and claims very sound.

Clarity: The text, especially the higher-level sections, is generally well-written and engaging, but as previously discussed the presentation of results is not as high quality.

Relation to Prior Work: The necessary citations are there, but sometimes the authors cite a long list of publications with insufficient context (L23, 85). These should be split into several chunks with distinctions/relevance clearly explained.

Reproducibility: Yes

Additional Feedback: I find your experiments extremely interesting with good results and insight, but unfortunately I don't think the current state of the presentation is publication-quality. I'd be interested in a comparison of your architecture findings to those of Ilyas et al.[1] In Figure 5, I would like to see intermediate (not after the final conv) layers included. Shape vs texture results would be interesting to contrast with the generic classification results of [2] when examining intermediate layers. Why did you opt for Exemplar as one of your methods in S6? It is fairly dated, especially if you're running it alongside of SimCLR. It seems to me like Instance Discrimination would be a better choice if you wanted a contrastive baseline to compare to SimCLR. The findings on BigBiGan are very interesting, but I think baselines of purely generative and encoder-based methods should be included as ablation (e.g. GAN and VAE) The findings of S7 Paragraph2 should be presented in a small table or figure rather than simply listing numbers. Similar comments for Paragraph3. Typo?: L254, "do not necessarily"? L307/8: Which experiments are you referring to as OOD? [1] Ilyas, Andrew, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. "Adversarial examples are not bugs, they are features." In Advances in Neural Information Processing Systems, pp. 125-136. 2019. [2] Asano, Yuki M., Christian Rupprecht, and Andrea Vedaldi. "A critical analysis of self-supervision, or what we can learn from a single image." arXiv preprint arXiv:1904.13132 (2019). UPDATE AFTER REBUTTAL: As the author's addressed my concerns concerning the visual presentation, I am happy to recommend acceptance (8) for this paper as well. The experiments and discussion are extremely well done and updated figures will make for an extremely complete and high quality work.

Review 2

Summary and Contributions: The paper presents an in-depth study of texture bias in CNNs. The authors use three datasets, the cue conflict dataset from Geirhos et. al. 2019, the Navon Dataset and ImageNet-C, and measure the effect of a variety of factors including the effect of data augmentation, self-supervised training objectives and model architecture. In my opinion there are two central findings: 1. CNNs can learn to classify based on shape equally well as on texture 2. Data augmentation during training plays a central role determining if a model is biased towards texture or shape. There is a wealth of additional findings including: - the effect of augmentations increasing the shape bias is cumulative - improving the shape bias is often at odds with ImageNet accuracy - self-supervised training does not by itself create models biased towards shape - shape bias somewhat correlates with accuracy in high-performing ImageNet models - shape information is mostly discarded in the last layers - innocent hyperparameters like random cropping, crop size and learning rate influence the bias and a number of more subtle points. Especially the experiment showing the additive effect of many data augmentation techniques and how they negatively influence ImageNet accuracy is extremely interesting. If a number of data augmentations can flip the model to having a shape bias but harms ImageNet accuracy that indicates that ImageNet may indeed not be suited to learn object recognition but rather encourages models to use texture and other subtle whole image cues that don’t require identifying the foreground. This seems to be very much in line with the results found in Kolesnikov et. al. 2019 Appendix G and if this effect is true it has the potential to shift the way the community vies ImageNet.

Strengths: Without a doubt this is the most in-depth study on shape and texture bias in CNNs and why they arise. There is a huge amount of insides in this paper and almost nothing that’s not to like.

Weaknesses: There are very few weaknesses, most of which I’d simply love to see because the work got me excited for the topic rather than things that are missing. The one thing that comes closes to an issue is with the self-supervised representation experiment specifically with the fine-tuning step. As the authors have established the shape bias of a model is strongly influenced by the data augmentation during training. However the data augmentation during fine-tuning is standard random flipping and cropping. As a result the model will have a similar bias as the standard model if we assume that the representation before the classifier contains both, texture and shape information, which is to be expected from the experiments in Section 8. In my opinion it would be interesting to either perform the fine-tuning with the same data augmentations as in the cumulative experiment or add the self-supervised representations to the decoding experiment in Section 8. To say it again there is nothing wrong with the current experiment and we definitely learn something from it: simply switching to self-supervised pre-training does not lead to models with a shape bias. There is however more insight to be gained from this. Regarding the things I’d love to see I think a study of newer models that were trained with different data augmentation strategies (e.g. AdvProp, mixup, AugMix, ...), newer models (FixResNet, EfficientNet, …) and probably most interesting additional data (BiT, ResNeXt-wsl, NoisyStudent, ...). But that is more of a fully-fledged follow-up than something missing from the submission.

Correctness: From what I can judge all claims and methods appear to be correct.

Clarity: I found the paper very clearly written and easy to understand.

Relation to Prior Work: Yes, the paper does a very good job at discussing and structuring prior work. The difference between the prior work and the presented study are absolutely clear.

Reproducibility: Yes

Additional Feedback: References: Geirhos et. al. 2019: ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness Kolesnikov et. al. 2019: Big Transfer (BiT): General Visual Representation Learning On the broader impact section: I miss any discussion of the negative impact this work can have or rather that some of the main findings imply. Most of the broader impact statement discusses how this study can help identifying and closing the gap between human and machine vision but as of now the main finding is that performance on ImageNet, the benchmark we as a field rely on most, is to some extend anticorrelated with a shape bias and thus anti-correlated with a strong bias of human perception. This is definitely worrisome, and I miss that aspect to be clearly stated in the broader impact section. Right now, it reads more like an extended outlook advertising all the positive future impact this work could have. I agree with this potential impact but at least for me listing only this is not what I expect from a broader impact statement. ### Post-rebuttal comment: I thank the authors for updating the broader impact statement. The proposed addition sounds good and fixes my issue. ### General Post-rebuttal comment: Overall the rebuttal and the other reviews did not change my assessment that this is a good paper and should be accepted. I think it is noteworthy that the authors promised to extend section 5 with results for all layers and generally implemented a number of suggestions all strengthening the paper.

Review 3

Summary and Contributions: This paper offers a more detailed inspection of the popular claim that "ImageNet-trained CNNs are biased towards texture" and studies the underlying reason of such observation, leading to the conclusion that the differences "arise not from differences in their internal workings, but from differences in the data that they see." Then the paper answers multiple related questions regarding this central conclusion, such as how different self-supervised loss (i.e. regularizations) play different roles in learning the bias, and how different augmentation strategies play different roles.

Strengths: The paper contributes to an important research question generally regarding the learning behavior of convolutional neural networks, it is built upon massive amount of work and answer many questions thoroughly.

Weaknesses: Some questions are not answered well. For example, the results in Section 8 do not look very convincing (See details later).

Correctness: I believe so.

Clarity: Yes.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback: Overall, this paper studies an important research question regarding the learning behavior of convolutional neural networks. The paper is very well written, seemingly involves a massive amount of wordload, and answers most of the questions clearly with evidence and offer conjectures of the unanswerable questions to guide future research. Despite the high quality, I noticed several drawbacks and suggest the authors to address them. Major issues: -. In the abstract, the paper says the differences "arise not from differences in their internal workings, but from differences in the data that they see", which seems to suggest that whether the model learns texture or shape primarily depends on the data seen, yet in the experiments, the authors demonstrate that, with more carefully designed regularizations (termed as "self-supervised losses" in the paper), the model can be pushed to focus more on the shape. This empirical observation seems to contradict with the main claim in the abstract since I suppose losses are one of the "internal workings" (or what does "internal workings" mean exactly?). I suggest the authors to revise corresponding texts to reflect this more accurately. -. In Section 8, I'm not sure why the authors only study three parts of the AlexNet and two parts of the ResNet, why not report Figure 5 for every single representations one can get from these networks (this study does not seem to introduce new computational loads). I believe it's fine if the authors cannot address this thoroughly and consider this section as an add-on to a good paper, but then I will recommend the authors to stay humble with the conclusions in this section, for example, to rephrase line 62-65 which highlights this section. Also, the authors wrote "suggesting that these models’ classification layers remove shape information", what does "remove" mean exactly? I suppose the last layer has no ability to "remove" or "alter" any preceding representations, it can only prefer one over the other. -. In addition to the dataset mentioned in Section 3, a recent manuscript [1] introduced a new concept of shape-texture dichotomy through the frequency spectrum. I wonder if this main arugments of this paper also follows this understanidng of shape vs. texture, why or why not. Minor issues: "Broader Impact" section should not be numbered. Overall, I think this is a good paper, but I still suggest the authors to address the above issues to further improve the paper. [1] "High-frequency Component Helps Explain the Generalization of Convolutional Neural Networks." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. ------------------ After reading the rebuttal, I appreciate the authors addressing some of my concerns, my ratings stay unchanged and congrats to overall positive ratings.

Review 4

Summary and Contributions: This paper scrutinizes the origin of texture bias in convolutional neural networks. To do so, it presents some hypotheses for the existence of bias such as network architecture, learning process, or the data. And the answer to each hypothesis is revealed by experiments. This study can be regarded as a good follow up study of former paper(Robert Geirhos et al, 2019), where CNN models are heavily biased towards the texture when trained on ImageNet. As the former finds the 'phenomenon', this study explores some clues that make it happens.

Strengths: As the existence of texture bias in the convolution neural network(CNN) has been known in the machine learning society from a recent study, there is a lot of doubt about why it exists, where it comes from, and how to deal with it. I believe that this is a good empirical study that is needed to understand the bias and answer to them. Also, this study presents how several methods(that are believed to enhance performances on classification tasks) are related to the texture or shape bias of the model. And it can be a good reference for machine learning researchers to guide their model to be more robust and human-like recognition.

Weaknesses: This study provides limited theoretical contributions and is mainly supported by empirical experiments. It would make the arguments more logical if some mathematical proofs are followed.

Correctness: This study has the good design of experiments for finding proofs for each hypothesis. No flaws found.

Clarity: This paper is well written and easy to read. It starts with several questions on texture bias and findings from experiments via hypothesis. Empirical studies have proper figures and tables that make readers understand well.

Relation to Prior Work: As mentioned above, this paper is an in-depth following study of the former where CNN models are biased to texture. And I believe there is no prior work similar to this study.

Reproducibility: Yes

Additional Feedback: ===================== Comments after the reponses of authors: I appreciate authors for all the sincere responses addressing reviewers' comments. However, some of my concerns(e.g. mathematical proofs) still remain the same. Therefore, my overall score would not be changed. ===================== Though this study focuses on the classification task, it can be extended to other tasks where CNN is utilized such as object detection or segmentation. By comparing results among different tasks, we would find whether the origin and prevalence of texture are task-specific or not.