Communications of the ACM
New Computer Program Aims to Teach Itself Everything About Anything
The Learning Everything about Anything (LEVAN) open-source program searches text and images online to study all variations of a concept and learn which specific terms are relevant.
Credit: literasura.blogspot.com
Computer scientists at the University of Washington and the Allen Institute for Artificial Intelligence in Seattle say they have developed the first fully automated computer program that pairs textual and visual data to learn visual concepts.
The Learning Everything about Anything (LEVAN) open-source program searches books and images online to study all variations of a concept, and offers results as a comprehensive, browsable list of images. By examining the content of the images and identifying characteristic patterns using object-recognition algorithms, LEVAN learns which terms are relevant.
LEVAN scans the text in millions of books available on Google Books to find every instance of a concept, then uses an algorithm to filter out words with no visual association. After determining which phrases are relevant, the program conducts a Web image search to find corresponding photos.
Dictionaries, encyclopedias, and other information resources are increasingly providing visual information because it is easier to comprehend and faster to browse, but these resources often must be manually curated, says Allen Institute research scientist Santosh Divvala. By contrast, Divvala says LEVAN requires no human supervision and automatically learns the visual knowledge for any concept.
LEVAN launched in March with only a few concepts and has now tagged more than 13 million images with 65,000 different phrases.
From UW News (WA)
View Full Article
Abstracts Copyright © 2014 Information Inc., Bethesda, Maryland, USA
No entries found