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.
Introduction
Deep neural networks (DNNs) are notoriously overconfident and fragile. While they perform exceptionally well on clean data, their performance often plummets when faced with "corruptions" such as motion blur, sensor noise, or adverse weather. Standard training-time augmentations like MixUp or CutMix attempt to mitigate this by exposing models to varied inputs, yet they typically maintain "perfect" one-hot labels even when the resulting image is so degraded that the object is barely recognizable. This creates a fundamental mismatch: the model is forced to be 100% certain about a partially visible object. This paper addresses this "overconfidence gap" by arguing that as the input signal degrades, the target label should also become proportionally less certain.
Methodology
The proposed framework, inspired by Generative Diffusion Models (GDMs), utilizes a process termed "mollification" consisting of two coupled components: input degradation and label smoothing.
Input and Label Mollification
During training, images are corrupted using two main modalities: Gaussian noising (following a variance-preserving schedule) and blurring (applying low-pass filters via a heat-equation-based schedule). Crucially, the intensity of label smoothing is coupled to the intensity of the input corruption.
As shown in Figure 1, as the input image $\mathbf{x}_t$ becomes more corrupted (increasing $t$), the target label $\mathbf{y}_t$ is smoothed by mixing it with a uniform distribution. For noise, the smoothing level is determined by the Signal-to-Noise Ratio (SNR). For blurring, which lacks a traditional SNR, the authors use an information-theoretic proxy: the PNG compression size of the image.
Figure 3 demonstrates that PNG size reduces linearly with the blurring schedule, providing a principled way to scale label smoothing. This approach is theoretically supported by Dirichlet distributions, which encourage the network to match its predicted confidence to the actual amount of degradation in the input.
Experimental Results
The researchers evaluated the method on standard benchmarks (CIFAR-10/100, TinyImageNet, ImageNet) and their corrupted (-C) variants containing 15 types of real-world distortions.
- Robustness Gains: On CIFAR-10-C, the error rate dropped from 12.1% to 8.0%. On CIFAR-100-C, it improved from 36.9% to 28.5%. On the large-scale ImageNet-C benchmark, mollification improved the mean error rate by approximately 5 percentage points.
- Calibration: The method significantly reduced the Expected Calibration Error (ECE), meaning the model's confidence scores became much more reliable indicators of actual accuracy.
- Efficiency: The method introduces negligible computational overhead and can be integrated into existing training pipelines with minimal code changes.
Discussion and Conclusion
A key insight of this work is that standard label smoothing is superior to simple "tempering" because it pushes probability density toward other classes, better representing the ambiguity of a corrupted image. Spectral analysis further reveals that noise and blur augmentations act as effective proxies for a wide range of other corruptions because they share similar frequency-domain characteristics.
While a slight "robustness tax" (a minor drop in clean accuracy, e.g., ~0.5% on ImageNet) was observed, the gains in out-of-distribution robustness are substantial. Future research may explore class-specific noise structures and the potential of this coupling to defend against adversarial attacks.
Summary updated