Build a Residual Module and a Residual Convolutional Network | by Carla Martins

As always, using Keras and Tensorflow!

In the last article, we learned about Residual Connections and their importance in neural networks to prevent gradient vanishing. In this article we we will learn how to build a Residual module and how to implement a full convolutional neural network with residual connections, as well as to compare the performance with a sequential convolutional neural network. Keep reading!

The first step in this article is to build a residual module. If haven’t read my past articles on convolutional networks, I strongly recommend you do so, and get familiar with convolutional layers, batch normalization, and activation functions.

Now that you have it all, let’s start building a Residual module. Remember the following image, where we have an input X, a first convolutional layer followed by an activation function, and a second convolutional layer without an activation function. In the next step, that input value X and the result of the two convolution layers are summed up, and followed by an activation function.