Example building block of a ResNet, consisting of three sequential convolutional layers. The input is a 128-channel activation map, which is passed through 64 1 × 1 convolutional filters. The filters extract a 64-channel feature map. These features, after applying a ReLU activation, are then passed through 64 3 × 3 convolutional filters. The purpose of the first layer is to compress the channels for the 3 × 3 convolutional layer, which results in less optimizable parameters. Then, the ReLU activation is applied again and the final 1 × 1 convolutional layer expands the number of channels back to 128. Finally, these outputs are summed with the inputs via a skip-connection and passed through a ReLU activation.