Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2022) · 2022

CNN Filter DB - An Empirical Investigation of Trained Convolutional Filters

Paul Gavrikov, Janis Keuper

Why this publication matters

Researchers often reuse a pretrained vision model without knowing much about the filters inside it. CNN Filter DB provides more than a billion filters for examining shared patterns, unusual behavior, and poorly functioning components. That evidence can inform how models are selected and reused for new applications.

Abstract

Currently, many theoretical as well as practically relevant questions towards the transferability and robustness of Convolutional Neural Networks (CNNs) remain unsolved. While ongoing research efforts are engaging these problems from various angles, in most computer vision related cases these approaches can be generalized to investigations of the effects of distribution shifts in image data. In this context, we propose to study the shifts in the learned weights of trained CNN models. Here we focus on the properties of the distributions of dominantly used 3 × 3 convolution filter kernels. We collected and publicly provide a dataset with over 1.4 billion filters from hundreds of trained CNNs, using a wide range of datasets, architectures, and vision tasks. In a first use case of the proposed dataset, we can show highly relevant properties of many publicly available pre-trained models for practical applications: I) We analyze distribution shifts (or the lack thereof) between trained filters along different axes of meta-parameters, like visual category of the dataset, task, architecture, or layer depth. Based on these results, we conclude that model pre-training can succeed on arbitrary datasets if they meet size and variance conditions. II) We show that many pre-trained models contain degenerated filters which make them less robust and less suitable for fine-tuning on target applications. Data & Project website: https://github.com/paulgavrikov/cnn-filter-db

Abstract source ↗

Figures

A collection of learned convolution kernels from CNN Filter DB. — **Figure 1.** First 3 × 3 filters extracted of each convolution layer in a ResNet-18 trained on CIFAR-10. The filters show a clear loss of diversity and increasing sparsity with depth. The colormap range is determined layer-wise by the absolute peak weight of all filters in that layer. View in source ↗

**Figure 2.** Comparison of layer entropy and sparsity of overparameterized, robust, and regular classification models. Outliers are hidden for clarity. View in source ↗

Cite this paper

Download .bib

@inproceedings{gavrikov2022cnnfilterdb16,
  title = {{CNN Filter DB - An Empirical Investigation of Trained Convolutional Filters}},
  author = {Paul Gavrikov and Janis Keuper},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2022)},
  year = {2022},
  url = {https://openaccess.thecvf.com/content/CVPR2022/papers/Gavrikov_CNN_Filter_DB_An_Empirical_Investigation_of_Trained_Convolutional_Filters_CVPR_2022_paper.pdf}
}

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