A convolutional network (CNN) stacks convolution, activation, and downsampling (pooling) layers to extract increasingly abstract features from images, from edges to objects, before a classification layer. It exploits the spatial structure of the data and uses far fewer parameters than an MLP on images.