Communications of the ACM
DARPA Offers Free Watson-Like Artificial Intelligence
The U.S. Defense Advanced Research Projects Agency has unveiled an open source version of IBM's Watson named DeepDive. The system was tested against humans performing identical tasks, and "DeepDive came out ahead or at least equaled the efforts of the humans," says University of Wisconsin-Madison (UWM) professor Shanan Peters, who supervised the testing. The testers were researchers who read technical journal articles and cataloged them by understanding their content.
Primary programmer and UWM professor Christopher Re says in contrast to Watson, "DeepDive's goal is to extract lots of structured data" from unstructured data sources via probability-based learning algorithms. DeepDive also uses open source tools such as MADlib, Impala, and such low-level techniques as Hogwild. "Underneath the covers, DeepDive is based on a probability model; this is a very principled, academic approach to build these systems, but the question for us was, 'Could it actually scale in practice?'"Re says. "Our biggest innovations in DeepDive have to do with giving it this ability to scale."
Re notes the researchers currently use a RISC processor, but "we're trying to make a compiler, and we think machine learning will let us make it much easier to program in the next generation of DeepDive."
From EE Times
View Full Article
Abstracts Copyright © 2014 Information Inc., Bethesda, Maryland, USA
No entries found