VANCOUVER, British Columbia, Dec. 8, 2003 -- Intel Corporation researchers have released software that allows developers to build computers that can "learn" from their experience, using data to proactively improve their own accuracy and the ease with which we use them. The announcement was made today at the opening of the Neural Information Processing Systems Conference (NIPS2003).

The software enables computers to estimate the likelihood that something will happen by calculating how often it occurred in the past. The software can be used to enhance a wide variety of interactive and industrial computer applications -- everything from culling through huge databases of gene studies to spot promising proteins for new drugs to email systems that create a model of a person's behavior to decide how best to manage newly arriving messages on its own. The software is available through Intel’s Open Source Machine Learning Library (OpenML), a toolbox of functions that helps researchers develop machine learning applications.

"Intel wants computers to be more proactive," said David Tennenhouse, vice president in Intel’s Corporate Technology Group and director of research. "To do this they need to be able to learn from their experiences with users and the world around them. Using new statistical methods to identify key patterns, these systems will start anticipating the needs of their users and pre-computing responses to the most likely questions so that the answers will be instantly available the moment they are needed. Combined with faster microprocessors, OpenML is certain to drive an explosion of machine learning-based applications such as toys that respond to a child’s movements and networks of wireless sensors that will enhance our safety, productivity and stewardship of our environment."

"At Concurrent Pharmaceuticals we integrate proprietary computational methods, medicinal chemistry and discovery biology to create novel drugs," said John Baldwin, president and chief scientific officer of Concurrent Pharmaceuticals. "Our approach creates massive amounts of computationally generated information that needs to be integrated seamlessly into the drug discovery process to be effective. The availability of new, more sophisticated tools for machine learning, such as OpenML, improves our ability to utilize our computational technologies in drug discovery and will help us achieve our goal of more rapidly discovering new therapeutics to improve human health."

Accelerating Research Into Machine Learning

OpenML is based on "Bayesian" mathematical principles which essentially are the idea that the probability of future events can be calculated by studying their prior frequency. Because Bayesian models are based on data collected from experience, the more data obtained the better the predictions, and if the data changes, the results correct themselves.

"With the release of OpenML, researchers worldwide will have access to a large, well-implemented, open-source toolbox of graphical model algorithms -- one that is fast and reliable," said Professor Michael I. Jordan, University of California at Berkeley.

Faster microprocessors and improvements in the graphical mathematical models at the heart of Bayesian methods are all enabling real-time computer "machine learning" algorithms to run on standard PCs. The computer machine learning library is designed to increase innovation in this field by providing source code for a wide range of machine learning functions. This open source software release includes C++ source code for all of the library’s functionality and a royalty-free redistribution license. The OpenML main Web site is located at http://www.intel.com/research/mrl/pnl.

In one example of Open ML’s usefulness, Intel researchers used it to help create an audio/video speech recognition system that enables computers to detect a speaker’s face with a video camera and track his mouth movements. Learning to "read lips" enables much more accurate speech recognition, particularly at airports, shopping malls and other noisy environments.

NIPS2003 is the premier scientific meeting on neural computation. Presentations cover a wide range of topics, including algorithms and architectures; applications; brain imaging; cognitive science and artificial intelligence; control and reinforcement learning; emerging technologies; learning theory; neuroscience; speech and signal processing; and visual processing.