Deep Residual Learning for Image Recognition

Residual learning framework to ease the training of deeper neural networks. Layers are reformulated as learning residual functions with reference to layer inputs, instead of learning unreferenced functions. These residual networks are easier to optimize and can gain accuracy from considerably increased depth. Deeper nets with a depth of up to 152 layers that are 8 times deeper than VGG nets are evaluated. 3.57% error rate on ImageNet is achieved. An 28% relative improvement on the COCO object detection dataset is achieved.