DARPA intends to revolutionize machine learning

Virtually every news item from DARPA, related to robots and artificial intelligence, is inevitably accompanied by scathing comments about Skynet. But this time they will be surprisingly appropriate. The Agency’s new research program is dedicated to probabilistic programming for solving advanced machine learning problems (Probabilistic Programming for Advanced Machine Learning or PPAML). According to program manager Kathleen Fisher, DARPA intends to “Do for machine learning what the emergence of high-level languages ​​50 years ago did for programming as a whole”.

Machine learning algorithms are already widely used in consumer technologies - anti-spam, speech recognition, auto-robots, and for analyzing huge amounts of data in medicine or finance. Naturally, the prospects of machine learning are also interesting for the military. At the same time, there are no generally accepted universal tools for creating intelligent systems. Because of this, it is necessary to constantly reinvent the wheel, to implement algorithms similar to two drops of water over and over again, to build an architecture from scratch.

The set of approaches and paradigms used in machine learning has been called probabilistic programming . Tools, libraries and programming languages ​​for him until they leave the walls of universities, and their list is quite short. DARPA intends to change this situation.

The program’s goals include a radical reduction in the complexity of creating machine learning systems, lowering the threshold for entering intelligent applications into programming, improving basic machine learning algorithms, maximizing the use of modern hardware technologies such as multi-core processors and GPUs, cloud computing, creating and standardizing API for connecting the elements of machine learning infrastructure into a single system.
')
The PPAML program is designed for 46 months. Detailed terms and conditions will be announced on April 10 at a presentation in Arlington, Virginia. So far, you can download a PDF with a brief description of the program.