Dealing with the Evil Twins: Improving Random Augmentation by Addressing Catastrophic Forgetting of Diverse Augmentations

2025/06/09 by Dongkyu Cho, Rumi Chunara, Cho, Dongkyu +1 · 1 citation
Computer Science · #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Face recognition and analysis #Generative Adversarial Networks and Image Synthesis #Machine Learning (cs.LG)

paper · pdf · doi:10.48550/arxiv.2506.08240

openalex publication_date 2025/06/09 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Data augmentation is a promising tool for enhancing out-of-distribution generalization, where the key is to produce diverse, challenging variations of the source domain via costly targeted augmentations that maximize its generalization effect. Conversely, random augmentation is inexpensive but is deemed suboptimal due to its limited effect. In this paper, we revisit random augmentation and explore methods to address its shortcomings. We show that the stochastic nature of random augmentation can produce a set of colliding augmentations that distorts the learned features, similar to catastrophic forgetting. We propose a simple solution that improves the generalization effect of random augmentation by addressing forgetting, which displays strong generalization performance across various single source domain generalization (sDG) benchmarks.

Citations

• PEER pressure: Model-to-Model Regularization for Single Source Domain Generalization
• Crafting Distribution Shifts for Validation and Training in Single Source Domain Generalization
• Continual Learning with Weight Interpolation
• AdvST: Revisiting Data Augmentations for Single Domain Generalization
• Weighted Ensemble Models Are Strong Continual Learners
• Domain Generalization by Rejecting Extreme Augmentations
• Out-of-distribution forgetting: vulnerability of continual learning to intra-class distribution shift
• A Comprehensive Survey of Continual Learning: Theory, Method and Application
• How Much Data Are Augmentations Worth? An Investigation into Scaling Laws, Invariance, and Implicit Regularization
• Diverse Weight Averaging for Out-of-Distribution Generalization
• FOSTER: Feature Boosting and Compression for Class-Incremental Learning
• Learning to Diversify for Single Domain Generalization
• Adversarially Adaptive Normalization for Single Domain Generalization
• Progressive Domain Expansion Network for Single Domain Generalization
• Maximum-Entropy Adversarial Data Augmentation for Improved\n Generalization and Robustness
• Just Pick a Sign: Optimizing Deep Multitask Models with Gradient Sign Dropout
• Domain Generalization for Medical Imaging Classification with Linear-Dependency Regularization
• Learning to Learn Single Domain Generalization
• Gradient Surgery for Multi-Task Learning
• RandAugment: Practical automated data augmentation with a reduced search\n space
• Data Augmentation Revisited: Rethinking the Distribution Gap between\n Clean and Augmented Data
• Similarity of Neural Network Representations Revisited
• Complementary Learning for Overcoming Catastrophic Forgetting Using\n Experience Replay
• Experience Replay for Continual Learning
• AutoAugment: Learning Augmentation Policies from Data
• Generalizing to Unseen Domains via Adversarial Data Augmentation
• Averaging Weights Leads to Wider Optima and Better Generalization
• Deeper, Broader and Artier Domain Generalization
• Deep Hashing Network for Unsupervised Domain Adaptation
• Gradient Episodic Memory for Continual Learning
• Learning Structured Sparsity in Deep Neural Networks
• Deep Residual Learning for Image Recognition
• Explaining and Harnessing Adversarial Examples
• Unsupervised Domain Adaptation by Backpropagation
• An Empirical Investigation of Catastrophic Forgetting in Gradient-Based Neural Networks
• The MNIST Database of Handwritten Digit Images for Machine Learning Research [Best of the Web]

Cited by

• Expert-guided Clinical Text Augmentation via Query-Based Model Collaboration

Related

• Untapped Potential of Data Augmentation: A Domain Generalization Viewpoint
• How Much Data Are Augmentations Worth? An Investigation into Scaling Laws, Invariance, and Implicit Regularization
• Optimizing Data Augmentation Policy Through Random Unidimensional Search
• Towards Understanding Why Data Augmentation Improves Generalization
• AugMax: Adversarial Composition of Random Augmentations for Robust Training