In case you've been living under a rock the past few weeks, Google's foray into artificial neural networks has yielded hundreds of thousands of phantasmagoric images. The company has an obvious interest in image classification, and here's how they explain the DeepDream process in their Research Blog:
Inceptionism: Going Deeper into Neural Networks
. . .
We train an artificial neural network by showing it millions of training examples [of dogs and eyes and pagodas, let's say] and gradually adjusting the network parameters until it gives the classifications we want. The network typically consists of 10-30 stacked layers of artificial neurons. Each image is fed into the input layer, which then talks to the next layer, until eventually the “output” layer is reached. The network’s “answer” comes from this final output layer.
. . .
One way to visualize what goes on is to turn the network upside down and ask it to enhance an input image in such a way as to elicit a particular interpretation. Say you want to know what sort of image would result in “Banana.” Start with an image full of random noise, then gradually tweak the image towards what the neural net considers a banana... By itself, that doesn’t work very well, but it does if we impose a prior constraint that the image should have similar statistics to natural images, such as neighboring pixels needing to be correlated.
After Google released the deepdream code on GitHub, Psychic VR Lab set up a Deep Dream web interface, which currently has over 300,000 groovy and scary images.
I've taken an interest in the hallucinogenic and distorted brain images, including the one above. I can't properly credit the human input interface (which wasn't me), but I found it after a submitting a file of my own in the early stages of http://psychic-vr-lab.com/deepdream/. I can't find the url hosting my image, but I came across the frightening brain here, along with the original.
I've included a few more for your viewing pleasure. Brain Decoder posted a dreamy mouse hippocampus Brainbow.
Here's one by HofmannsBicycle.
And a fun fave courtesy of @rogierK and @katestorrs. This one is cartoonish instead of menacing.
@sarcastic_f this is the JHU atlas dreamified pic.twitter.com/wQ1dbZeLgJ
— rogier kievit (@rogierK) July 18, 2015
Rogier said: "According to #deepdream the homunculus in our brains is a terrifying bird-dog hybrid."
Aw, I thought it was kind of cute. More small birds, fewer staring judgmental eyeballs.
And the grand finale isn't a brain at all. But who doesn't want to see the dreamified version of The Garden of Earthly Delights, by Hieronymus Bosch? Here it is, via @aut0mata. Click on image for a larger view.
When nothing's right, just close your eyes
Close your eyes and you're gone
-Beck, Dreams
ADDENDUM (July 21 2015): It's worth reading Deepdream: Avoiding Kitsch by Josh Nimoy, which confirms the training set was filled with dogs, birds, and pagodas. Nimoy also shows deepdream images done with neural networks trained on other datasets.
For example, the image below was generated by a neural network trained to do gender classification.
Gendernet deepdreaming Cindy Sherman Untitled B (1975)
Google AI is apparently a dog person.
ReplyDeleteBut seriously, do many of the dremified images have critters in them it makes me wonder about the distribution of the image type of the training sets.
I think you're right. All dogs and birds. I saw some foxes early on, but never any cats. Guess they didn't do any cat/dog training (gif).
ReplyDeleteOh I stand corrected! If you zoom in to 200-300% on the Bosch picture, there's one creature that looks a bit like a lion. There's also a rabbit/dog hybrid where it's hard to see them both at the same time.
ReplyDeleteAnd that may be a fish on someone's head.
That's pretty interesting, but I must say that I prefer the original Bosch. In the Netherlands, you can buy tangible objects taken out of paintings. I have two of the eggmen from the Garden. I couldn't figure out a way to attach photos here.
ReplyDeleteRight Mind Matters - Sounds pretty cool!
ReplyDeleteAnonymous - A post by Josh Nimoy confirms the training set was filled with dogs, birds, and pagodas.