Keuper Labs
← All publications

Transactions on Machine Learning Research (TMLR) · 2024

Learning the essential in less than 2k additional weights - a simple approach to improve image classification stability under corruptions

Kai Bäuerle, Patrick Müller, Syed Muhammad Kazim, Ivo Ihrke, Margret Keuper

RobustnessImage ClassificationSignal ProcessingEfficient Deep Learning

Why this publication matters

A small addition to the input pipeline can make an existing classifier more stable under image corruption. This work learns that addition using fewer than 2,000 extra weights. It demonstrates that improving reliability can sometimes begin with a lightweight preprocessing step rather than a larger network.

Abstract

The performance of image classification on well-known benchmarks such as ImageNet is remarkable, but in safety-critical situations, the accuracy often drops significantly under adverse conditions. To counteract these performance drops, we propose a very simple modification to the models: we pre-pend a single, dimension preserving convolutional layer with a large linear kernel whose purpose it is to extract the information that is essential for image classification. We show that our simple modification can increase the robustness against common corruptions significantly, especially for corruptions of high severity. We demonstrate the impact of our channel-specific layers on ImageNet-100 and ImageNette classification tasks and show an increase of up to 30% accuracy on corrupted data in the top1 accuracy. Further, we conduct a set of designed experiments to qualify the conditions for our findings. Our main result is that a data- and network-dependent linear subspace carries the most important classification information (the essential), which our proposed pre-processing layer approximately identifies for most corruptions, and at very low cost.

Abstract source ↗

Figures

Classification accuracy across optical corruptions with and without the preprocessing layer.
Figure 1. A pre-pended large kernel convolution layer (ours) can increase the robustness of classification networks against unknown corruptions without additional data augmentation. Here, we show ResNet50 improved with a trainable pre-pended filter evaluated on ImageNette (Howard, 2023) blur and corruptions from OpticsBench (Müller et al., 2023). For each corruption type, five levels of severity are shown from left to right. The variation, visualized via the box plots, results from five different seeds per model. View in source ↗
The architecture of our proposed trainable input layer.
Figure 2. The architecture of our proposed trainable input layer. We learn a single depth-wise convolution to encourage the model to represent the essential, i.e. the part of the input data that is crucial for classification. View in source ↗

Cite this paper

Download .bib 
@article{bauerle2024learningtheessential51,
  title = {{Learning the essential in less than 2k additional weights - a simple approach to improve image classification stability under corruptions}},
  author = {Kai Bäuerle and Patrick Müller and Syed Muhammad Kazim and Ivo Ihrke and Margret Keuper},
  journal = {Transactions on Machine Learning Research},
  year = {2024},
  url = {https://openreview.net/pdf?id=i2SuGWtIIm}
}

Figures and abstract are reproduced from the linked research sources. Credit remains with the authors and publishers.