Computer Learns Common Sense By Looking At Pictures All Day; Potential Step Toward Strong Artificial Intelligence
Researchers at Carnegie Mellon University have created a machine that can look at pictures, recognize patterns, and learn the common-sense relationships between different objects. They call it NEIL — the Never Ending Image Learner — and it's been learning images around the clock since July, the university announced this week.
Like Us on Facebook
Already, NEIL has analyzed more than 3 million images, collected from Google images searches. It has learned, for example, that deer look like antelope, that cars have wheels, and that the leaning tower is in Pisa. The goal, the scientists say, is to expand a field of study around computer vision, in which machines recognize colors and shapes. "What we have learned in the last five to 10 years of computer vision research is that the more data you have, the better computer vision becomes," said Abhinav Gupta, the lead author of the research, in the statement.
An enormous amount of data, they say, is what allows NEIL to be so remarkable. By processing thousands of images of the same things, say, an Airbus 330 and an airplane nose, it begins to recognize the characteristics of each and connect the dots. In this way, they learn common sense in much the same way humans do: by seeing the world. "Images are the best way to learn visual properties," says Gupta, an assistant research professor in Carnegie Mellon's Robotics Institute. "Images also include a lot of common sense information about the world. People learn this by themselves and, with NEIL, we hope that computers will do so as well."
The machine "is a computer program that runs 24 hours per day and seven days per week to automatically extract visual knowledge from Internet data," the team writes on its website. "It is an effort to build the world's largest visual knowledge base with minimum human labeling effort — one that would be useful to many computer vision and [artificial intelligence] efforts." NEIL so far has learned about 2,500 associations.
And whereas other computer vision efforts have involved assistance from people in identifying pictures, this project is almost exclusively automatic. The scientists must only get involved for two reasons: One, they have to tell the computer what objects or topics to look for in the first place, and, two, they have to help suss out confusing homonyms. Apple, for example, can be a fruit and a technology company, and poor NEIL doesn't know the difference until they show it.
Gupta and two doctoral students are expected to present on their project, funded by the Defense Department's Office of Naval Research and Google, next week at the IEEE International Conference on Computer Vision in Sydney, Australia. While they're there, NEIL's hundreds of processors will continue to rifle through image data. Gupta says they have no plans to terminate the experiment; they're curious to see how much it can possibly learn on its own.
© 2012 iScience Times All rights reserved. Do not reproduce without permission.