DeepDream sending a picture via the network and then figuring out how gradient it is in relation to a certain layer's activations. Then, by altering the image to amplify these activations, the network is able to perceive more patterns, producing a visual that resembles a dream.