Google researchers have fed images into the tech giant’s artificial neural network and the image recognition software has output some fairly trippy images reminiscent of psychedelic hallucinations.
In order to produce the images, the researchers asked Google’s image recognition neural network to recognize a feature — such as buildings, animals and objects — and modify the picture in such a way as to emphasize the feature it recognizes. The modified output was then fed back into the system in a loop which ran until the output had been altered beyond recognition, The Guardian reported.
Shontee Pant, a staff writer with the Christian Science Monitor, referred to the images created by the neural network at the request of the researchers as “eerie and beautiful, sometimes both at the same time.”
The experiment was conducted by Alexander Mordvintsev, Christopher Olah, and Mike Tyka.
Neural networks, also known as neural nets, are computer systems modeled after the human brain and nervous system. In essence, such networks are a system of programs and data structure which approximate the operation of the human brain. Generally, neural networks involve a large number of processors operating in parallel with one another, each within its own small sphere of knowledge with access limited to its local memory.
In the experiment, researchers fed an image full of random noise into the network and gradually tweaked it towards what the system considers a banana. While the initial output indicated that the experiment “didn’t work very well,” according to the researchers, instilling a prior constraint that the image should have similar statistics to natural images resulted in an interesting image depicting blurry banana-esque objects in a glitchy fashion.
In other news, a DARPA-funded program has programmed a robot which learns to cook by watching how-to videos on YouTube.
