# Deep Residual Learning for Image Recognition

**Authors:** Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun  
**arXiv:** [1512.03385](https://arxiv.org/abs/1512.03385)  
**Date:** 2015-12-10  
**Summary page:** https://www.summarizepaper.com/en/paper/1512.03385v1/

## Key points

- Authors: Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun
- Introduced a novel residual learning framework for training deeper neural networks
- Redefined layers as learning residual functions in relation to layer inputs
- Empirical evidence shows easier optimization and higher accuracy with increased depth
- Extensive experiments on ImageNet dataset with depths up to 152 layers
- Achieved 3.57% error rate on ImageNet test set, securing first place in ILSVRC 2015 classification task
- Benefits of deep representations demonstrated on CIFAR-10 with 100 and 1000 layers
- Played foundational role in successful submissions to ILSVRC & COCO 2015 competitions
- Secured first place in tasks such as ImageNet detection, localization, COCO detection, and segmentation
- Remarkable 28% relative improvement on COCO object detection dataset due to extremely deep representations

## Summary

In their paper titled "Deep Residual Learning for Image Recognition," authors Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun address the challenge of training deeper neural networks. They introduce a novel residual learning framework that facilitates the training of networks with significantly greater depth than previously utilized models. By redefining the layers as learning residual functions in relation to the layer inputs, rather than learning unreferenced functions, they demonstrate through empirical evidence that these residual networks are easier to optimize and can achieve higher accuracy with increased depth. The authors conducted extensive experiments on the ImageNet dataset, evaluating residual nets with depths of up to 152 layers - surpassing VGG nets by eight times while maintaining lower complexity. Their ensemble of residual nets achieved an impressive 3.57% error rate on the ImageNet test set, securing first place in the ILSVRC 2015 classification task. Additionally, they analyzed performance on CIFAR-10 using 100 and 1000 layers, further showcasing the benefits of deep representations in visual recognition tasks. Notably, their deep residual nets played a foundational role in their successful submissions to the ILSVRC & COCO 2015 competitions. They secured first place in tasks such as ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation - underscoring the significance of depth in representations for achieving superior results in object detection and segmentation tasks. Furthermore, due to their extremely deep representations, the authors observed a remarkable 28% relative improvement on the COCO object detection dataset. This highlights how leveraging deep residual networks can lead to significant advancements in visual recognition tasks and underscores their effectiveness in enhancing model performance across various datasets and competitions.