Google Research Blog:
Google software engineers have taken these image-detection algorithms and tweaked the process, so it essentially goes in reverse. Instead of recognizing existing features in an image, the computer looks for and generates imagery where there isn't any already.
The results of this process are psychedelic, slightly surreal artificial visions that the Google engineers call "Inceptionism".
For instance, the network can be given an image and instructed to search for edges, which are enhanced and filtered with colors.
It can also "dream" of images that aren't there, by searching for familiar forms and reinforcing them- as in this image of a cloudy sky, in which the computer pulls out animal-like forms. This is similar to a human daydreaming and finding recognizable shapes in the clouds.
The important aspects of this process are the image recognition feedback loops, applied to the visual analysis that the computer is making. It is an iterative process of searching for recognizable features, then re-applying it over and over, to create hallucinatory fractal visuals.
The brain normally suppresses neural activity and feedback patterns that would be hallucinatory & distracting, but I suspect substances that unlock visionary states open up this activity.
Although this may help us to understand what is going on when a person has a psychedelic experience, the research is also useful for understanding consciousness, in general.
For more images from Google's neural net research, you can visit the photo gallery of Google software engineer Michael Tyka:
There is even an interactive livestream Twitch channel of a neural net bot that creates ever-changing psychedelic imagery in real time.
The visual output is based on user suggestions, derived from a predetermined list of inputs. These inputs are categories of a million images that were previously introduced to the network.
Watch live video from 317070 on www.twitch.tv