The Mystery of Google Deep Dream

Recently, posts of some strange pictures have been showing up on my Facebook and Twitter feed. The images seem to be from the drunken mind of Salvador Dali…

Sadly, they are not, but it turns out they are just as interesting.

After some research, it turns out that these images come from an artificial intelligence program invented  by Google called “Deep Dream.”  To put it bluntly, Google’s Deep Dream is probably one of the trippiest computer programs you’re likely to see.  All you have to do is upload an image, tell the program to look for patterns, and wait as Google reveals exactly what your picture looks like to it. The pattern recognition employed here is not unlike what the human mind creates using its own pattern recognition faculties. 

The whole Deep Dream program started when Google decided to train an artificial neural network by showing it millions of training examples and gradually adjusting network parameters until the network gave the classifications it wanted.  In this case, Google aimed to create a program that could search for images.  Needless to say, the results turned out to be both more interesting and more complex. 

Yet one of the challenges of neural networks is understanding what exactly goes on at each layers of the artificial neurons. Each image is fed into the input layer, when then talks to the next layer until the “output” layer is reached.  What’s fascinating, though, is that this program can actually show researchers what it’s “thinking” and “seeing.”  These terms, of course, are defined loosely and only in the context of what data a computer program is outputting. 

As an example, I decided to submit my headshot to Deep Dream and see what would happen.  Below is the result: 

The Deep Dream Program relies on the idea that the program recognizes shapes and patterns and then generates images on the picture.  It essentially works to recognize objects within a picture and then overlays the objects on the picture. 

Why do the pictures look so trippy? That’s because Deep Dream actually undergoes something similar to what humans experience during hallucinations. It’s been given free rein to follow the impulse of any recognizable imagery and exaggerate it in a self-referencing loop.  Again, not unlike what the human mind experiences while dreaming or under the influence of hallucinogens such as Tryptamine. 

Google has actually open sourced the code that they used to generate the strange-looking image. In fact, you can download it all for yourself here.

With that said, there are also several websites that have popped up where you can use this code. Just upload your image and wait for the time it takes to process.  This time has been a few days for me, but others report it can take up to a week.  If you do upload a photo, I recommend keeping the file size as small as possible without losing the detail of the image itself (about 5mb in my experience).

In my own experiments, I have submitted several different photos.  Most recently, I was delighted at the result when I decided to post a photo of a painting I completed a few weeks ago that currently hangs in my dining room.  The painting, entitled “The Dance of Dragons,” can be seen here below in its completely unaltered form:

Unaltered Photo of “The Dance of Dragons”

When I put this photo into the Deep Dream program, the results were pretty staggering, as I believe some of you may agree.  See below for the results (You can click on either image to zoom in):

“The Dance of Dragons” after a trip through Google Deep Dream

Projects such as this one give me great hope for the future.  I believe that with continued research, we will soon begin to fully understand the nature of the most powerful tool Nature has given us.  Our minds. 

In the mean time, I’m happy to find ways to manipulate thoughts in the context of entertainment, and in the meantime look at some pretty trippy images from one of the most interesting artificial intelligence programs I have seen.  If anyone that’s reading this decides to submit an image and gets some really interesting results, please post them below and I would be happy to feature them on the site.