Core Concepts
The author proposes Auxiliary Fourier-basis Augmentation (AFA) as a technique to enhance model robustness in image classification by filling the gap left by visual augmentations. AFA uses additive noise based on Fourier-basis functions efficiently and seamlessly integrates with other augmentation techniques.
Abstract
The paper introduces Auxiliary Fourier-basis Augmentation (AFA) to address the limitations of common visual augmentations in improving model robustness in real-world scenarios. AFA targets augmentation in the frequency domain, complementing existing techniques and demonstrating enhanced performance against common corruptions, OOD generalization, and consistency of predictions under perturbations. By utilizing Fourier-basis functions for additive noise, AFA offers an efficient approach that bridges the gap left by traditional visual augmentations. The method is shown to be computationally efficient, allowing for training larger models on larger datasets while maintaining or even improving generalization results. Through a combination of main and auxiliary components, AFA ensures robustness against adversarial distribution shifts induced by frequency-based noise. The study includes experiments on benchmark datasets like CIFAR-10, CIFAR-100, TinyImageNet, and ImageNet, showcasing the effectiveness of AFA in enhancing model performance across various metrics.
Stats
Models trained with AFA showed improved standard accuracy (SA) and robust accuracy (RA).
AFA demonstrated reduced mean corruption error (mCE) across different datasets.
The proposed method contributed to better generalization performance on benchmark datasets.
Ablation analysis highlighted the importance of auxiliary components in improving model robustness.
Comparison between ACE loss and JSD loss showed minimal differences in robustness performance.
Sensitivity analysis of hyperparameter 1/λ indicated low sensitivity to its choice.
Quotes
"Auxiliary Fourier-basis Augmentation (AFA) benefits the robustness of models against common corruptions, OOD generalization, and consistency of predictions w.r.t. perturbations."
"AFA efficiently bridges the gap left by traditional visual augmentations through additive noise based on Fourier-basis functions."