Researchers: Computer teaches itself everything about anything
SEATTLE - We've all been there, sitting in front of our Internet device, searching for answers to our questions, navigating the idiosyncrasies of search tools, Google, Bing, Yahoo, and all the others. Two Seattle computer scientists believe they developed a technology that can narrow down Internet search results faster than the search engine behemoths, all using a computer that teaches itself about concepts.
They call it LEVAN, short for Learning Everything about Anything. It's a cloud-based computer program the researchers claim can teach itself everything there is to know about a visual concept, and make the results searchable for anyone to access.
"No one has done this in the world, and we believe it will have a lot of impact for users who want to know more about concepts," says Santosh Divvala, a computer scientist at the Allen Institute for Artificial Intelligence. Divvala along with Ali Farhadi, an associate professor of Computer Science at the University of Washington began collaborating on LEVAN two years ago.
LEVAN operates differently than traditional search engines, because the researchers say it's design to learn to understand the concepts behind a specific image. The patent pending technology searches text from millions of books written in English, and available on Google Books, then an algorithm filters out words that aren't visual.
Once it has learned which phrases are relevant, LEVAN does an image search on the Internet looking for uniformity in appearance among the photos its retrieved. Its looking for a common relationship in each pixel of each relevant picture, and if there's a match, LEVAN connects the image to the text phrases.
The result is a series of common images that can be connected to a search query.
For example, a user has an image of a specific chair, but doesn't know what type of chair it is. Typing in the word "chair" in Google, gives the user a long list of results involving chairs, but the results may not be related to each other. A search of Google Images results in an unrelated mixture of pictures as well.
The researchers say LEVAN produces images of chairs categorized by their types and styles because LEVAN's technology has learned the difference between images of all the chairs on the Internet.
Right now, LEVAN is limited in how fast it can learn about a concept because the computational power to process a query takes up to 10 hours. Currently it has 3,000 concepts in the cue, and another 10,000 waiting in the wings.
Interested in learning more? Give LEVAN a try.