See Convolutional neural network, Deep art
Basically, we feed the image forward, set the gradients equal to the activations, and then backpropagate to the image. Update the image along the calculated gradient and repeat.
This amplifies any features that the network slightly detected. Kind of like Apophenia