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.
|Task||Dataset||Model||Metric name||Metric value||Global rank||Compare|
|Image Classification||CIFAR-10||ResNet-1001||Percentage correct||95.38||# 10|
|Image Classification||CIFAR-10||ResNet-1001||Percentage error||4.62||# 10|
|Image Classification||CIFAR-100||ResNet-1001||Percentage correct||77.28999999999999||# 5|
|Image Classification||CIFAR-100||ResNet-1001||Percentage error||22.71||# 5|