# Aggregated Residual Transformations for Deep Neural Networks

**Authors:** Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He  
**arXiv:** [1611.05431](https://arxiv.org/abs/1611.05431)  
**Date:** 2016-11-16  
**Summary page:** https://www.summarizepaper.com/en/paper/1611.05431v1/

## Key points

- Authors introduced a novel network architecture for image classification
- The network is highly modularized and constructed by repeating a building block that aggregates transformations with the same topology
- Introduction of "cardinality" as a key concept, referring to the size of the set of transformations in the network
- Increasing cardinality demonstrated to significantly improve classification accuracy on ImageNet-1K dataset
- Increasing cardinality more effective in enhancing performance than increasing depth or width of the network
- ResNeXt models developed by authors achieved 2nd place ranking in ILSVRC 2016 classification task
- Superior results on ImageNet-5K set and COCO detection set compared to ResNet counterpart
- Highlighted cardinality as crucial alongside depth and width when designing deep neural networks for image classification tasks

## Summary

In their paper titled "Aggregated Residual Transformations for Deep Neural Networks," authors Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He introduce a novel network architecture for image classification. The proposed network is highly modularized and constructed by repeating a building block that aggregates a set of transformations with the same topology. This design results in a homogeneous, multi-branch architecture with minimal hyper-parameters to set. One key innovation introduced by the authors is the concept of "cardinality," which refers to the size of the set of transformations in the network. They demonstrate empirically on the ImageNet-1K dataset that increasing cardinality can significantly improve classification accuracy even when maintaining complexity constraints. Notably, increasing cardinality proves more effective in enhancing performance than simply increasing depth or width of the network. The models developed by the authors - known as ResNeXt - served as the foundation for their participation in the ILSVRC 2016 classification task where they achieved an impressive 2nd place ranking. Furthermore, their investigation extended to an ImageNet-5K set and the COCO detection set, showcasing superior results compared to its ResNet counterpart. Overall, this work highlights cardinality as a crucial factor alongside depth and width when designing deep neural networks for image classification tasks. The success of ResNeXt models underscores potential improvements through thoughtful architectural choices and parameter settings.