DARPA launches PPAML artificial intelligence program to move machine learning forward

ARLINGTON, Va., 26 March 2013. U.S. military researchers are kicking off an artificial intelligence program to make machine learning capability much more accessible and effective for a wide variety of military weapons and information systems.

Scientists at the U.S. Defense Advanced Research Projects Agency (DARPA) in Arlington, Va. will brief industry on 10 April 2013 on the Probabilistic Programming for Advancing Machine Learning (PPAML), which seeks rapid advancements in machine learning, which experts say is at the heart of modern approaches to artificial intelligence.

The PPAML industry briefings will be from 10 a.m. to 2 p.m. at the Executive Conference Center at Liberty Center, 4075 Wilson Blvd., Suite 350, in Arlington, Va.

The goal of PPAML is to advance machine learning by using probabilistic programming to increase the number of people who can build machine learning applications; make machine learning experts more effective; and enable new applications that are impossible to conceive of using today’s technology.

Specifically, the PPAML program will try to make machine learning model code shorter; reduce development time; help build richer models; reduce the expertise necessary build machine learning applications; and support the construction of integrated models.

Machine learning assumes that enabling computers to learn can be more effective than programming them explicitly -- especially in military applications such as intelligence, surveillance, and reconnaissance (ISR); natural language processing (NLP); predictive analytics; cyber; and various scientific disciplines.

Example of machine learning include self-driving cars, image search and activity detection, object tracking, topic models, spam filters, recommender systems, predictive databases, and gene sequencing.

Although enabling computers to learn from their experiences can be valuable, building effective machine learning applications today requires Herculean efforts on the part of trained experts in machine learning, DARPA researchers point out.

Clearing this hurdle will research breakthroughs in several areas. For the front end of the system, the primary issues boil down to simplifying language while getting information efficiently; making the system usable by many people; devising profilers, debuggers, and model verification/checking tools to determine accuracy.

For the back end of the system, the key research challenge is improving system performance and predictability. To do this, DARPA researchers want to develop analyses that select the most appropriate solver for a particular query; improve the performance of existing solvers by blending in ideas from the compiler optimization community; compiling specific solvers for multi-core machines, graphics processing units (GPUs), cloud infrastructures, and custom hardware; developing new solvers; and developing an API for new solvers.

Research challenges in basic machine learning technology will include advancing the theory of probabilistic programming; discovering new efficient inference algorithms; discovering efficient representations; developing inference algorithms that work over streaming data; and developing techniques for assessing model fitness.

The PPAML industry briefings are to familiarize participants with DARPA’s approaches to machine learning by using probabilistic programming; identify potential proposers; and promote teaming. A formal solicitation for the PPAML program should be released by the end of this month.

Companies interested should register for the PPAML briefings no later than 5 April online at www.solers.com/BAAinfo-reg/ppaml. For questions or comments email DARPA at [email protected].

More information is online at https://www.fbo.gov/spg/ODA/DARPA/CMO/DARPA-SN-13-30/listing.htm.