PushPull-Conv: Improving the robustness of traditional convolutional neural networks (ConvNets) against typical image corruptions, by a simple architectural change (biologically inspired). Replace the first Conv layer with the PushPull-Conv unit to achieve robustness. The goal is to achieve robustness to certain unseen distributions during model learning.
Datasets used: CIFAR-10, ImageNet-100, ImageNet-200, and ImageNet-1k.
Models were trained on the clean images, and tested on corrupted versions (CIFAR10-C and ImageNet-C; hendrycks et. al).
An example of the 15 types of corruption with 5 levels of severity is shown below:
Results on ImageNet are presented below. For other datasets and models please refer to the paper. In the Table below, E is the top1 clean error, mCE is mean corruption error and R_net is net reduction in error rate when compared to the baseline. R_net is expressed as a percentage, and accounts for both E and mCE (Refer to Sec. 5.2 in the paper for further details).
| Variants of baseline models | E ↓ | mCE↓ | R_net↑ |
|---|---|---|---|
| ResNet50 (baseline) | 0.269 | 0.667 | 0.00 |
| ResNet50 + PushPull avg3 (PP3) | 0.282 | 0.645 | 1.67 |
| ResNet50 + PushPull avg5 (PP5) | 0.276 | 0.645 | 2.79 |
| ResNet50 + AutoAug | 0.269 | 0.630 | 6.86 |
| ResNet50 + AutoAug + PP3 | 0.283 | 0.604 | 9.05 |
| ResNet50 + AugMix | 0.269 | 0.612 | 10.14 |
| ResNet50 + AugMix + PP3 | 0.287 | 0.590 | 10.97 |
| ResNet50 + PRIME | 0.298 | 0.522 | 21.57 |
| ResNet50 + PRIME + PP3 | 0.306 | 0.500 | 24.31 |
- The current implementation is suited to run on HPC Clusters, an example script is located here.
- A simplified script is being developed to quickly run both the training and the predict flows.
If you find this useful in your work, please consider citing:
@misc{bennabhaktula2024push,
title={{PushPull-Net: Inhibition-driven ResNet} robust to image corruptions},
author={Guru Swaroop Bennabhaktula and Enrique Alegre and Nicola Strisciuglio and George Azzopardi},
year={2024},
eprint={2408.04077},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2408.04077},
}