In the last few years, CNNs have become popular in the areas of image recognition, object detection, segmentation, and many other tasks in the field of computer vision. They are also becoming popular in the field of natural language processing (NLP), though they are not commonly used yet. The fundamental difference between fully connected layers and convolution layers is the way the weights are connected to each other in the intermediate layers. Let's take a look at an image where we depict how fully connected, or linear, layers work:
One of the biggest challenges of using a linear layer or fully connected layers for computer vision is that they lose all spatial information, and the complexity in terms of the number of weights used by fully connected layers is too big. For example, when we represent a 224 pixel image as a flat array, we would...
