ResNet (Deep Residual Learning for Image Recognition)

Status: Read

Author: Kaiming He, Xiangyu Zhang

Topic: CNNs, CV , Image

Category: Architecture

Conference: CVPR

Year: 2016

Link: https://openaccess.thecvf.com/content_cvpr_2016/html/He_Deep_Residual_Learning_CVPR_2016_paper.html

Summary: Introduces Residual or Skip Connections to allow increase in the depth of a DNN

Questions

What did authors try to accomplish?

Authors try to increase the depth of a DNN while resolving the issues of vanishing gradients due to depth.

What were the key elements of the approach?

Introduces Residual / Skip connections between layers to propagate gradients from higher layers directly to lower layers.

What can you use yourself from this paper?

Residual connections generally give better performance in many different models especially when the model is deep.
Such connections are also made between encoder-decoder models to propagate features learned from earlier layers directly to higher layers. Eg. see U-Net model.

What other references to follow?