Androids Dreaming of Electric Sheep
But I digress. Neural nets are good at the more human tasks; recognising faces, or understanding human speech. Google Images uses a neural net to be able to search for pictures of things based on the words you type in. They have indexed millions of images, tagging them with words, like “red”, “dog”, “picnic” and the neural net begins to make it’s own connections between the patterns in images that are all tagged with “dog”, and so it eventually learns pretty well what a dog looks like. After that the neural net doesn’t need humans to tell it there is a dog in a certain picture, it can figure it out for itself. So you can go to images.google.com and type in “dog” and it gives you a bunch of pictures that have dogs in them. Now, some creative engineers at Google recently came up with the idea of having a look inside their vast image neural net, by asking it to output a picture of what it is seeing after having begun to apply its recognition process to an image.
recently came up with the idea of having a look inside their vast image neural net, by asking it to output a picture of what it is seeing after having begun to apply its recognition process to an image.
It’s like showing a picture to your friend, then asking them to draw what they saw; the picture they draw for you reveals some of the original picture, and some of the thoughts that arose in your friend’s mind. It tells you a bit about how they think. + The images produced by this process are best described as psychedelic… a stream of computerised consciousness, “I see an eye, I think I see a dog, I see a dog, I see an eye, colours”.
These images are being tagged with the hashtag #deepdream.
(The title for this piece is a play on the title of the science fiction novel, “Do Androids Dream of Electric Sheep?” by Philip K Dick; nothing to do with Android phones.) I have massively simplified the technical specifics of this process, for brevity, clarity, and because, having failed neural nets as a subject, I’m not totally up on those specifics. I
recommend this article if you are interested in more indepth information :
This photo of our backyard is transformed into a far stranger place by the dreams of the Google neural net.