Identity Mappings in Deep Residual Networks

About

Deep residual networks have emerged as a family of extremely deep architectures showing compelling accuracy and nice convergence behaviors. In this paper, we analyze the propagation formulations behind the residual building blocks, which suggest that the forward and backward signals can be directly propagated from one block to any other block, when using identity mappings as the skip connections and after-addition activation. A series of ablation experiments support the importance of these identity mappings. This motivates us to propose a new residual unit, which makes training easier and improves generalization. We report improved results using a 1001-layer ResNet on CIFAR-10 (4.62% error) and CIFAR-100, and a 200-layer ResNet on ImageNet. Code is available at: https://github.com/KaimingHe/resnet-1k-layers

Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun · 2016

Related benchmarks

Task	Dataset	Result
Image Classification	CIFAR-100 (test)	Accuracy75.67	3,518
Image Classification	CIFAR-10 (test)	Accuracy94.54	3,381
Object Detection	COCO 2017 (val)	AP38.5	2,930
Image Classification	ImageNet-1k (val)	Top-1 Accuracy—	1,498
Semantic segmentation	Cityscapes (test)	mIoU—	1,254
Image Classification	ImageNet (val)	Top-1 Acc23.77	1,206
Image Classification	CIFAR-10 (test)	Accuracy—	906
Image Classification	MNIST (test)	Accuracy—	894
Image Classification	CIFAR-100 (val)	Accuracy—	781
Image Classification	CIFAR-100	Accuracy—	691