Challenges of image recognition with fully connected networks

In Chapter 5, Image Classification with Neural Networks, we applied a deep neural network (DNN) to the Fashion MNIST dataset. We saw how every neuron in the input layer is connected to every neuron in the hidden layer and those in the hidden layer are connected to neurons in the output layer, hence the name fully connected. While this architecture can solve many ML problems, they are not well suited for modeling image classification tasks, due to the spatial nature of image data. Let’s say you are looking at a picture of a face; the positioning and orientation of the features on the face enable you to know it is a human face even when you just focus on a specific feature, such as the eyes. Instinctively, you know it’s a face by virtue of the spatial relationship between the features of the face; however, DNNs do not see this bigger picture when looking at images. They process each pixel in the image as independent...