# RegNet: Self-Regulated Network for Image Classification

**Authors:** Jing Xu, Yu Pan, Xinglin Pan, Steven Hoi, Zhang Yi, Zenglin Xu  
**arXiv:** [2101.00590](https://arxiv.org/abs/2101.00590)  
**Date:** 2021-01-03  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2101.00590v1/

## Key points

- ResNet and its variants have achieved remarkable success in computer vision tasks
- Shortcut connection mechanism used in these networks limits their ability to explore new complementary features
- Authors propose a regulator module as a memory mechanism to extract complementary features and enhance ResNet performance
- Regulator module is composed of convolutional RNNs (Convolutional LSTMs or Convolutional GRUs)
- New architecture named RegNet, regulator module can be easily implemented and appended to any ResNet architecture
- Regulator module applied to improve Squeeze-and-Excitation ResNet, demonstrating generalization ability
- Experimental results on three image classification datasets show that RegNet outperforms standard ResNet and other state-of-the-art architectures like SE-ResNet
- Incorporating regulator module enhances performance of ResNet in extracting complementary features and improving overall classification accuracy
- RegNet introduced as a self-regulated network for image classification tasks, achieves promising results compared to existing architectures

## Summary

. The ResNet and its variants have achieved remarkable success in various computer vision tasks. However, the simple shortcut connection mechanism used in these networks limits their ability to explore new complementary features. To address this issue, the authors propose a regulator module as a memory mechanism to extract complementary features and enhance the performance of ResNet. The regulator module is composed of convolutional RNNs, such as Convolutional LSTMs or Convolutional GRUs, which are known for their ability to extract spatio-temporal information. The authors name the new architecture RegNet. The regulator module can be easily implemented and appended to any ResNet architecture. Additionally, the authors apply the regulator module to improve the Squeeze-and-Excitation ResNet, demonstrating the generalization ability of their method. Experimental results on three image classification datasets show that RegNet outperforms both the standard ResNet and other state-of-the-art architectures like SE-ResNet. This indicates that incorporating the regulator module enhances the performance of ResNet in extracting complementary features and improving overall classification accuracy. Overall, this paper introduces RegNet as a self-regulated network for image classification tasks. By incorporating a regulator module based on convolutional RNNs, RegNet overcomes limitations in exploring new potentially complementary features and achieves promising results compared to existing architectures.