The complex process of image recognition is one of the busiest areas of research in artificial intelligence. Labs around the world are trying to teach machines to actually see the world around them, processing still images and video in such a way that they can comprehend - or at least accurately describe - what they're looking at.
Developments on this front are slowly trickling down into practical applications. Facebook is working on a neural network solution to captioning images and videos for the blind. And Google recently cut loose its Deep Dream system that lets machines apply a kind of trippy lateral thinking process to image recognition.
Putting the Art in Artificial Intelligence: Photos
Then there's this intriguing video currently making the round, in which U.S. artist and coder Kyle McDonald essentially takes an A.I. for a walk through Amsterdam to find out what the program actually "sees."
Specifically, McDonald loaded up his laptop with the open source program NeuralTalk, a system developed at Stanford University that applies natural language descriptions to any image it analyzes. Point your camera at any given scene and the system attempts to describe what it sees by comparing the image - at lightning-fast speeds - with millions of images it has previously cataloged.
McDonald modified the NeuralTalk program so that it would analyze footage from his webcam, on the fly, as he wandered around the city. It's a fascinating thing to watch as the machine attempts to describe the constantly changing surroundings in natural language captions.
The program is pretty good with the straightforward stuff: "A city street with a lot of car and a bus." But it's also easily confused. An image of a boat generates the caption: "A computer keyboard sitting on top of a wooden table." But then the system squints: "A boat is parked on the side of the river."
Facial Recognition System Detects Pain
It's amazing to see how much the neural network gets right, and how fast it can think. But it's also a little spooky to see how often the system is off - really off. Science fiction writers have been riffing darkly on this premise for decades. Machines like RoboCop and the Terminator have a dubious track record with image recognition and targeting.
Anyway, I'm sure it's nothing to worry about. You can check out the video below.
via The Verge