Components of convolutional neural networks – Towards Data Science

Recent state-of-the-art architectures have employed a number of additional components to complement the convolution operation. In this post, I would be explaining some of the most important components that have improved both the speed and accuracy of modern convolutional neural networks. I would begin by explaining the theory of each of the components and finalize with a practical implementation in keras. The first secret sauce that has made CNNs very effective is pooling. Pooling is a vector to scalar transformation that operates on each local region of an image, just like convolutions do, however, unlike convolutions, they do not have filters and do not compute dot products with the local region, instead they compute the average of the pixels in the region (Average Pooling) or simply picks the pixel with the highest intensity and discards the rest (Max Pooling).