Robust Classification by Coupling Data Mollification with Label Smoothing

TL;DR

The paper addresses robustness of image classification to test-time corruptions by coupling input mollification (noising and blurring) with label smoothing. It frames mollification within a probabilistic augmentation perspective, introducing an augmented likelihood and practical schedules for degrading inputs and labels over a unit interval $t \in [0,1]$. Empirically, the approach yields improved robustness and uncertainty calibration across CIFAR-10/100, TinyImageNet, and ImageNet-C, with notable gains in corrupted-image accuracy and reliable calibration, while incurring minimal overhead. The work also provides insights into when to favor noise versus blur, offers a Dirichlet interpretation of smoothing, and outlines directions for extending mollification to adversarial and class-specific noise settings, highlighting its practical relevance for improving real-world robustness.

Abstract

Introducing training-time augmentations is a key technique to enhance generalization and prepare deep neural networks against test-time corruptions. Inspired by the success of generative diffusion models, we propose a novel approach of coupling data mollification, in the form of image noising and blurring, with label smoothing to align predicted label confidences with image degradation. The method is simple to implement, introduces negligible overheads, and can be combined with existing augmentations. We demonstrate improved robustness and uncertainty quantification on the corrupted image benchmarks of CIFAR, TinyImageNet and ImageNet datasets.

Figures (11)

• Figure 1: Mollification augments training with perturbed images (a) and smoothed labels (c). Input mollification and label smoothing follow monotonic schedules (d), which reflect the signal-to-noise ratio of the images (b). Label smoothing prefers predictions whose distribution matches label uncertainty (e). The method extends to an arbitrary number of classes.
• Figure 2: The logarithmic heat blurring schedule on a TinyImageNet 64$\times$64 image.
• Figure 3: Blur reduces information linearly.
• Figure 4: Label smoothing is an intuitive way to degrade multi-class labels. The Dirichlet $\mathrm{Dir}(\mathbf{f} | \mathbf{1} + \mathbf{y})$ visualizations show cross-entropy $\mathbf{y}^\mathrm{onehot}$ (a) and label tempering $\mathbf{y}^\mathrm{temp}$ (b) to retain prediction mode at $(1,0,0)$, while label smoothing $\mathbf{y}^\mathrm{LS}$ (c) prefers non-peaky predictions.
• Figure 5: Adding mollification ($\cdots$) to augmentations ($\cdots$) improves corrupted accuracy (a), likelihood (b) and calibration (c) over CIFAR-10, CIFAR-100 and TinyImageNet.