Robust Classification by Coupling Data Mollification with Label Smoothing

Markus Heinonen, Ba-Hien Tran, Michael Kampffmeyer, Maurizio Filippone

Introduction

Deep neural networks (DNNs) have achieved remarkable success in image classification, yet they remain notoriously "brittle." While a model might achieve near-perfect accuracy on a clean validation set, its performance often collapses when faced with real-world corruptions such as motion blur, rain, or sensor noise. A significant factor in this fragility is the mismatch between the model's training objective and the reality of corrupted data.

Standard training typically employs "one-hot" labels, forcing the model to be 100% confident in a single class regardless of image quality. When data augmentation is used to introduce noise or blur, the model is still pushed toward extreme confidence on these degraded inputs. This leads to poor calibration, where the model's reported confidence does not reflect its actual accuracy on ambiguous or corrupted data.

The Data Mollification Framework

Inspired by the success of generative diffusion models (GDMs), this paper proposes a training strategy called data mollification. In GDMs, data is gradually transformed into pure noise through a controlled process. The authors adapt this concept to classification by coupling input degradation with adaptive label smoothing. The core philosophy is simple: as the input signal becomes less certain due to corruption, the target label should become less certain as well.

1. Input Mollification

During training, images are randomly corrupted using two primary modalities:

  • Noising: Adding Gaussian noise according to a variance-preserving schedule.
  • Blurring: Applying low-pass filters derived from the heat equation, implemented efficiently via Discrete Cosine Transforms (DCT).

2. Label Mollification (The "Coupling")

The innovation lies in tying the intensity of these corruptions to the target label distribution. Instead of static one-hot vectors, the authors use Label Smoothing (LS), where the target is a mixture of the ground-truth class and a uniform distribution over all classes.

Figure 2: Overview of the coupling mechanism. As the input image x_t becomes noisier or blurrier (a), the target label y_t (c) transitions from a one-hot vector toward a uniform distribution, guided by the schedules in (b) and (d).

As shown in the figure above, if an image is clean ($t=0$), the label remains nearly one-hot. As the corruption level increases ($t \to 1$), the label becomes increasingly uniform. This prevents the network from learning to be overconfident on inputs where the identifying features have been obscured.

Theoretical Basis: The Dirichlet Interpretation

The authors justify the use of label smoothing over other techniques (like temperature scaling) through a Dirichlet interpretation. By viewing the target distribution as the mean of a Dirichlet distribution, they demonstrate how label smoothing effectively pulls the target away from the extreme corners of the probability simplex.

Figure 3: Comparison of target distributions on the probability simplex. Label smoothing (c) provides a more robust target for corrupted data compared to standard cross-entropy (a) or simple tempering (b).

This geometric shift ensures that the model learns a more nuanced representation of uncertainty. By matching predicted confidence to the amount of noise in the input, the model becomes more "honest" about its limitations, significantly improving uncertainty calibration.

Key Results and Robustness

The method was evaluated on major benchmarks including CIFAR-10/100, TinyImageNet, and ImageNet, specifically focusing on the "-C" (corrupted) variants which contain 15 types of unseen real-world corruptions.

  • CIFAR-10-C: The corrupted error rate was reduced from 12.1% (using standard TrivialAugment) to 8.0%.
  • CIFAR-100-C: Error rates dropped from 36.9% to 28.5%.
  • ImageNet-C: The method provided significant gains across nearly all corruption types, particularly in noise and blur categories, with negligible impact on clean validation accuracy.

A key insight is the spectral similarity between the simple Gaussian noise/heat-blur used in training and complex real-world corruptions like snow or JPEG artifacts. Because these share similar frequency distributions, training on mollified data generalizes remarkably well to diverse test-time distortions.

Conclusion

By coupling input mollification with adaptive label smoothing, this framework bridges the gap between high accuracy and reliable calibration. Forcing models to acknowledge the loss of information in corrupted signals leads to more robust feature representations. Future work may explore class-specific noise structures and the interaction between this approach and adversarial robustness.

Summary updated

BibTeX
@inproceedings{heinonen2025robust,
  author = {Markus Heinonen and Ba-Hien Tran and Michael Kampffmeyer and Maurizio Filippone},
  title = {{Robust Classification by Coupling Data Mollification with Label Smoothing}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2025},
  url = {https://openreview.net/forum?id=YfI2zJDMid},
}
Figure