Making the “dream” images is very simple. Essentially it is just a gradient ascent process that tries to maximize the L2 norm of activations of a particular DNN layer.
A DNN is a Deep Neural Network, a computer simulation of a biological neural network with many layers.
Each layer consists of an array of simulated neurons, and every input to the network activates a specific set of neurons in each layer.