Google's Deepdream

Nillbugwtw

Zombier than thou.
Community Admin
Donor
Joined
Aug 6, 2011
Messages
959
Reaction score
1,894
Points
243
What are deep dream images?
Deep Learning is a new field within Machine Learning. In the past 4 years researchers have been training neural networks with a very large number of layers. Algorithms are learning how to classify images to a much greater accuracy than before: you can give them an image of a cat or a dog and they will be able to tell the difference. Traditionally this has been nearly impossible for computers but easy for humans. Deep Learning algorithms are trained by giving them a huge number of images, and telling them what object is in each image. Once it has seen (e.g.) a hundred types of dog heads 1000 times from a hundred angles, it has been 'trained'. Now you can give it new images and it will spot dog heads within the images, or tell you that there are none at all.

These things are... sorta disturbing. They present a world full of dogslugs and demons.





 

Lee_scar

Geezuslike
Donor
Joined
Aug 8, 2011
Messages
715
Reaction score
596
Points
63
i've read the paragraph in the OP but what made the pictures in the spoiler
What you do is you run an image thru the algorithm that detects dog heads and such and it prints out the computers visualization rather its inspection and essentially you are seeing what deep dream see's and understands from a photo before spitting out its result. very nightmarish

There are a few links in froyds post but I have to go and tag them nsfw and not responsible for content because I just came from there and there was a tad bit of not all ages rated content.
 

Lee_scar

Geezuslike
Donor
Joined
Aug 8, 2011
Messages
715
Reaction score
596
Points
63
You tell the computer to find dogs in the image and it goes nuts. It's what a neural network's dreams might look like.
I just thought in retrospect that by telling it to find dogs in the image, we are telling it that dogs exist in a photo where a dog may or may not exist. So in essence we are causing a huge margin of error and disrupting the program. No?
 

Pick Yer Poison

is a kid now, is a squid now
Joined
Aug 5, 2011
Messages
616
Reaction score
1,731
Points
243
Website
itisntrocketscience.tumblr.com
I just thought in retrospect that by telling it to find dogs in the image, we are telling it that dogs exist in a photo where a dog may or may not exist. So in essence we are causing a huge margin of error and disrupting the program. No?
Nope! This is run on a separate section of the program. It attempts to find dogs using what it expects a dog to look like, but does not add the data to its set of learning data. We are utilizing the data, but not adding to it.