# Global Pooling, More than Meets the Eye: Position Information is Encoded Channel-Wise in CNNs

**Authors:** Md Amirul Islam, Matthew Kowal, Sen Jia, Konstantinos G. Derpanis, Neil D. B. Bruce  
**arXiv:** [2108.07884](https://arxiv.org/abs/2108.07884)  
**Date:** 2021-08-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.07884v1/

## Key points

- Global pooling in CNNs does not eliminate all spatial information, but encodes positional information based on the ordering of channel dimensions
- The authors propose a data augmentation strategy and loss function to enhance translation invariance in CNNs, improving their ability to handle variations in object position and orientation
- They introduce a method for efficiently determining which channels encode overall position information or region-specific positions in the latent representation of a CNN
- Semantic segmentation heavily relies on overall position channels for accurate predictions
- It is possible to perform a "region-specific" attack by degrading a network's performance in specific parts of an input
- This work challenges conventional assumptions about global pooling and opens up new avenues for improving translation invariance and exploring targeted attacks within CNN architectures.

## Summary

In their paper titled "Global Pooling, More than Meets the Eye: Position Information is Encoded Channel-Wise in CNNs," authors Md Amirul Islam, Matthew Kowal, Sen Jia, Konstantinos G. Derpanis, and Neil D. B. Bruce challenge the widely held assumption that global pooling eliminates all spatial information when collapsing the spatial dimensions of a 3D tensor in a convolutional neural network (CNN) into a vector. They demonstrate that while semantic information is largely not preserved, positional information is encoded based on the ordering of the channel dimensions. Building upon this insight, the authors present two practical applications to showcase the real-world implications of their findings. Firstly, they propose a simple yet effective data augmentation strategy and loss function that enhances the translation invariance of a CNN's output. This approach improves the network's ability to handle variations in object position and orientation. Secondly, they introduce a method for efficiently determining which channels in the latent representation of a CNN are responsible for encoding overall position information or region-specific positions. Through experiments, they reveal that semantic segmentation heavily relies on overall position channels to make accurate predictions. Furthermore, they demonstrate for the first time that it is possible to perform a "region-specific" attack by degrading a network's performance in specific parts of an input. The authors believe that their findings and demonstrated applications will greatly benefit research areas focused on understanding the characteristics of CNNs. By challenging conventional assumptions about global pooling and uncovering how positional information is encoded channel-wise, this work opens up new avenues for improving translation invariance and exploring targeted attacks within CNN architectures.