Adversarial AutoMixup
Huafeng Qin, Xin Jin, Yun Jiang, Mounim A. El-Yacoubi, Xinbo Gao
TL;DR
AdAutomixup introduces an adversarial data augmentation method that generates challenging mixed samples to improve the robustness and accuracy of image classifiers, outperforming existing mixup techniques.
Contribution
It proposes an adversarial approach to generate hard mixed examples, enhancing model robustness and diversity in data augmentation for image classification.
Findings
Outperforms state-of-the-art methods on seven image benchmarks.
Generates challenging samples that improve classifier robustness.
Effective in various classification scenarios.
Abstract
Data mixing augmentation has been widely applied to improve the generalization ability of deep neural networks. Recently, offline data mixing augmentation, e.g. handcrafted and saliency information-based mixup, has been gradually replaced by automatic mixing approaches. Through minimizing two sub-tasks, namely, mixed sample generation and mixup classification in an end-to-end way, AutoMix significantly improves accuracy on image classification tasks. However, as the optimization objective is consistent for the two sub-tasks, this approach is prone to generating consistent instead of diverse mixed samples, which results in overfitting for target task training. In this paper, we propose AdAutomixup, an adversarial automatic mixup augmentation approach that generates challenging samples to train a robust classifier for image classification, by alternatively optimizing the classifier and…
Peer Reviews
Decision·ICLR 2024 spotlight
* (**S1**) This paper provides an interesting view of improving mixed sample qualities through adversarial training in the close-loop optimized mixup augmentation framework. The overall presentation of the manuscript is easy to follow, and the proposed methods are well-motivated. * (**S2**) Extensive experiments on mixup benchmarks verify the performance gains of the proposed AdAutoMix compared to existing mixup methods. Popular Transformer architectures are included in experiments.
* (**W1**) More empirical analysis of the proposed methods can be added. Despite the authors visualizing the mixed samples and CAM maps of various mixup methods, it can only reflect the overall performances and characteristics of different methods. I suggest the authors provide a fine-grained analysis of each proposed module to demonstrate its effectiveness, e.g., plotting the classification accuracy of using adversarial training or not. * (**W2**) Small-scale experiments. The authors only prov
Results are consistently better than AutoMixup and the evaluation (Table 1) is thorough. -------- Post-rebuttal: The authors have adequately addressed the concerns in the review. Useful experiments and ablations have been added as well. I'm still a little skeptical about the actual impact of the paper, from the methods and corresponding evaluation numbers in the paper I believe that we're at the point of diminishing returns. I've therefore increased my score to a 6.
There is no evaluation compared to Adversarial data augmentation approaches [1, 2, 3, 4]. At least an introduction or related works section should be added as relevant approaches to the problem. The term “cross attention module” (CAM) should not be used as it can be confused with “class activation mapping” (CAM) which is generally used in saliency-based data augmentation methods. Some notation is confusing - the encoder weight is updated with an EMA of the weights of the classifier - $\hat{\p
The idea to augment with hard examples is interesting. Furthermore, to iterate between augmentation and classifier is also interesting. Showed results are strong.
I do not see any significant weakness. The method is harder to implement and it requires more resources that other augmentation techniques, but given the timeline of augmentation, it is expected
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdversarial Robustness in Machine Learning · Advanced Neural Network Applications · Domain Adaptation and Few-Shot Learning
MethodsMixup