Hands-On Java Deep Learning for Computer Vision

ResNet-50

One of the successful architectures that uses the residual networks is ResNet-50, and it took first place in 2015 on the ImageNet dataset, with 1,000 classes. The error rate was quite low, that is, 3.57%. It was a simple architecture, because it uses the convolution same, like VGG-16, and max pooling. It applies these previous activations to deeper layers and then repeats this for every layer. It has only one different trick that we didn't mention previously: once you use the pooling layer, in contrast to the convolution same, the pooling layer doesn't preserve the first two dimensions but shrinks them, by dividing by two. That causes an obstacle, because you can't simply forward this activation to the further activation layers; ...