[CentralOH] Google's Deep Dreams Python code
eric at intellovations.com
Fri Aug 14 15:22:57 CEST 2015
A few months ago, Google posted a blog entry describing their image
classification artificial neural network (ANN) and how they wanted to
visualize better what was going on in each layer of the 10-30 layer ANN.
They realized that their ANN is also actually pretty good at *generating*
images as well. One way is to take a random image and tweak it with some
constraints along a path that results in the ANN considering it closer to
what it has been trained to recognize.
Another way is to take an image and ask different layers of the ANN to
enhance the features it is recognizing. Basically asking the ANN, "Whatever
you see here, give me more of it". So you end up with some pretty awesome
art. There is a link to a gallery on the blog post.
Google has open-sourced the code, and provided an IPython notebook showing
how they do it here. It's a very well documented notebook:
There are already apps using this code, just search for "Google Deep Dream".
Pretty cool stuff!
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the CentralOH