In this example, we will look at a slightly different generative network. We will see how to take a pre-trained convolutional network and use it to generate new objects in an image. Networks trained to discriminate between images learn enough about the images to generate them as well. This was first demonstrated by Alexander Mordvintsev of Google and described in this Google Research blog post (https://research.googleblog.com/2015/06/inceptionism-going-deeper-into-neural.html). It was originally called inceptionalism but the term deep dreaming became more popular to describe the technique.
Deep dreaming takes the backpropagated gradient activations and adds it back to the image, running the same process over ...
