Thursday, January 8, 2015

Deep Learning: How to Teach Your Robot Iron Chef

From Singularity Hub:
If robots and AI are our technological children (and of course they are!), what's the best way to teach them about the world? Why, the internet, of course. Using the popular deep learning programming technique, computer scientists are rearing the next generation of infant AIs on a steady diet of online images and videos.
A bit like the scene in the Matrix when Neo downloads kung fu directly to his brain, deep learning programs rapidly absorb large amounts of data and learn from it. Of course, the former is fictional—the latter, not at all.

In computer vision, for example, programs fed thousands of images can independently learn to isolate and identify individual components in them. While you won't just stumble upon ten thousand cat pictures in a desk drawer, the web is a treasure trove of such data. Fertile ground for young, impressionable programs.

Now, artificial intelligence researchers are moving beyond still images. University of Maryland researchers, for example, recently trained deep learning software with 88 YouTube cooking videos. After binging on the videos, the software learned to identify simple culinary tasks and form commands for a robot arm.

How does it work?

The program isolates hands in the video and assigns one of six “grasp” positions. It identifies objects and classifies them as one of 48 foods or tools. Finally, it identifies the action being performed—combining the lot into a command for execution by a robotic arm and grasper.

Simple as the tasks are, it isn’t easy teaching software from raw YouTube footage. Background variation and noise make it harder to pick out the video's critical elements. To further improve accuracy, the program calculates the most probable action by associating verbs and nouns in the video.

Put to the test, the deep learning software was able to accurately recognize and correctly classify objects and use what it saw in the videos to form commands for a robotic arm in various related actions. Looking forward, the team thinks software like theirs has great potential for robot learning....MORE
The program identifies hands and classifies their position.
The program identifies hand positions.
Related: