Cue-Timing Curriculum Learning for Robust CNN Image Classification
DOI:
https://doi.org/10.59992/IJCI.2026.v5n9p2الكلمات المفتاحية:
Corruption Robustness، Data Augmentation، Curriculum Learning، Shortcut Learning، Convolutional Neural Networksالملخص
Convolutional networks are brittle to common image corruptions, losing accuracy precipitously. We hypothesize this is because they rely on shortcut cues like texture and high-frequency detail. Critical-period theory predicts that learning such cues is an early event in training, from which there should be little recovery if they are suppressed early. We operationalize this prediction by scheduling per-cue suppression probabilities over training and comparing each schedule against an exposure-matched control that preserves overall suppression but randomizes its timing. Surprisingly, across 15 ResNet-18 trained on CIFAR-10, suppressing texture/high-frequency cues reduced mean corruption error by approximately ⅙ relative to ERM but concentrating that suppression budget early in training captured none of the benefit and was uniformly outperformed by its own control. Robustness tightly covaried with residual high-frequency sensitivity, which was reduced by 10x by uniformly distributing its exposure. It is the amount of exposure to cue suppression, not its schedule, that affects robustness.
المراجع
• Achille, A., Rovere, M., & Soatto, S. (2019). Critical learning periods in deep neural networks. arXiv. https://doi.org/10.48550/arXiv.1711.08856
• Bengio, Y., Louradour, J., Collobert, R., & Weston, J. (2009). Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning (pp. 41-48). ACM. https://doi.org/10.1145/1553374.1553380
• Cubuk, E. D., Zoph, B., Mane, D., Vasudevan, V., & Le, Q. V. (2019). AutoAugment: Learning augmentation strategies from data. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 113-123). IEEE. https://doi.org/10.1109/cvpr.2019.00020
• Cubuk, E. D., Zoph, B., Shlens, J., & Le, Q. V. (2020). RandAugment: Practical automated data augmentation with a reduced search space. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (pp. 3008-3017). IEEE. https://doi.org/10.1109/cvprw50498.2020.00359
• DeVries, T., & Taylor, G. W. (2017). Improved regularization of convolutional neural networks with Cutout. arXiv. https://doi.org/10.48550/arXiv.1708.04552
• Fadhilah, H., et al. (2026). Enhanced robustness in image classification through DistortionMix: A hybrid distortion-based augmentation technique. Jurnal Ilmu Komputer dan Informasi, 19(1). https://doi.org/10.21609/jiki.v19i1.1558
• Frankle, J., Schwab, D. J., & Morcos, A. S. (2020). The early phase of neural network training. arXiv. https://doi.org/10.48550/arXiv.2002.10365
• Fukase, V. Y., et al. (2025). One period to rule them all: Identifying critical learning periods in deep networks. Procedia Computer Science. https://doi.org/10.1016/j.procs.2025.07.138
• Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F. A., & Brendel, W. (2019). ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv. https://doi.org/10.48550/arXiv.1811.12231
• Geirhos, R., Jacobsen, J.-H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., & Wichmann, F. A. (2020). Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11), 665-673. https://doi.org/10.1038/s42256-020-00257-z
• He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (pp. 770-778). IEEE. https://doi.org/10.1109/cvpr.2016.90
• Hendrycks, D., & Dietterich, T. (2019). Benchmarking neural network robustness to common corruptions and perturbations. arXiv. https://doi.org/10.48550/arXiv.1903.12261
• Hendrycks, D., Mu, N., Cubuk, E. D., Zoph, B., Gilmer, J., & Lakshminarayanan, B. (2020). AugMix: A simple data processing method to improve robustness and uncertainty. arXiv. https://doi.org/10.48550/arXiv.1912.02781
• Hendrycks, D., Zou, A., Mazeika, M., Tang, L., Li, B., Song, D., & Steinhardt, J. (2022). PixMix: Dreamlike pictures comprehensively improve safety measures. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 16762-16771). IEEE. https://doi.org/10.1109/cvpr52688.2022.01628
• Hou, C., Zhang, J., & Zhou, T. (2023). When to learn what: Model-adaptive data augmentation curriculum. arXiv. https://doi.org/10.48550/arXiv.2309.04747
• Hwang, H., et al. (2024). Improving model robustness with frequency component modification and mixing. IEEE Access, 12. https://doi.org/10.1109/access.2024.3510923
• Kleinman, M., Achille, A., & Soatto, S. (2023). Critical learning periods for multisensory integration in deep networks. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 24296-24305). IEEE. https://doi.org/10.1109/cvpr52729.2023.02327
• Mittal, T., et al. (2026). Adaptive test-time augmentation via KL-regularized reinforcement learning for robust visual inference. Scientific Reports. https://doi.org/10.1038/s41598-026-62059-4
• Nam, J.-H., et al. (2024). FSDA: Frequency re-scaling in data augmentation for corruption-robust image classification. Pattern Recognition, 150, 110332. https://doi.org/10.1016/j.patcog.2024.110332
• Ranabhat, R. N., et al. (2025). Multi-scale unrectified push-pull with channel attention for enhanced corruption robustness. Proceedings of the AAAI Symposium Series, 6(1). https://doi.org/10.1609/aaaiss.v6i1.36022
• Siedel, G., et al. (2026). Stylized synthetic augmentation further improves corruption robustness. In Proceedings of the 21st International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS. https://doi.org/10.5220/0014251600004084
• Soviany, P., Ionescu, R. T., Rota, P., & Sebe, N. (2022). Curriculum learning: A survey. International Journal of Computer Vision, 130(6), 1526-1565. https://doi.org/10.1007/s11263-022-01611-x
• Wang, X., Chen, Y., & Zhu, W. (2021). A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9), 4555-4576. https://doi.org/10.1109/tpami.2021.3069908
• Yun, S., Han, D., Chun, S., Oh, S. J., Yoo, Y., & Choe, J. (2019). CutMix: Regularization strategy to train strong classifiers with localizable features. In 2019 IEEE/CVF International Conference on Computer Vision (pp. 6022-6031). IEEE. https://doi.org/10.1109/iccv.2019.00612.
• Zhang, H., Cisse, M., Dauphin, Y. N., & Lopez-Paz, D. (2018). mixup: Beyond empirical risk minimization. arXiv. https://doi.org/10.48550/arXiv.1710.09412.
• Zhang, Q. (2025). A two-phase data augmentation strategy with mixup and AugMix. In 2025 International Conference on Computers, Information Processing and Advanced Education (pp. 214-219). IEEE. https://doi.org/10.1109/cipae66821.2025.00041.
