The Pentagon Wants a Computer That Can Teach Itself

In an effort to pick up the pace of research and development in AI, the U.S. military’s advanced concepts research wing is launching an initiative to design automated tools that will make it easier to not just program computers, but to help computers teach themselves.

DARPA is not happy with the way machine learning is being done today — the practice of getting computers to act without having to be explicitly programmed. Which is not to say that machine learning hasn’t yielded tremendous benefits, which it most certainly has. Over the past decade, these systems have assisted in the development of self-driving cars, practical speech recognition, more powerful web searches, and email spam filters.

Moreover, machine learning has allowed developers to make machines more human-like in terms of their decision-making capacities, despite relying on techniques that use algorithms instead of neurological processes.

But the problem, says DARPA, is that each new application requires a “Herculean” effort to develop, even when performed by teams of specially-trained machine learning experts. A fundamental limitation is the lack of tools to build these systems.

Looking to overcome these problems, DARPA announced a new program to fund research in probabilistic programming languages.

These systems can comb through hideously large amounts of uncertain information and select the most useful parts of it. And unlike a traditional program, it can move backwards and forwards through the data, make inferences, reason under uncertainty, and efficiently home in on the best explanations. It can go through these processes in an iterative manner, repeating the exercise in a more efficient and superior way each time. Essentially, a probabilistic program learns and evolves.

To get the ball rolling, DARPA is holding a Proposer’s Day conference on Probabilistic Programming for Advanced Machine Learning this coming April 10th.

“We want to do for machine learning what the advent of high-level program languages 50 years ago did for the software development community as a whole,” said DARPA’s Kathleen Fisher through a statement.

Ultimately, DARPA hopes to foster the development of new applications that are outside the reach of today’s technology. In addition, it hopes to dramatically increase the number of people who can successfully build machine learning applications, while making machine learning experts “radically” more effective.

“Our goal is that future machine learning projects won’t require people to know everything about both the domain of interest and machine learning to build useful machine learning applications,” said Fisher. “Through new probabilistic programming languages specifically tailored to probabilistic inference, we hope to decisively reduce the current barriers to machine learning and foster a boom in innovation, productivity and effectiveness.”

The program is scheduled to last 48 months, so it will be some time before we see the benefits. But once realized, these scaled-up machines will be able to perform a number of complex tasks, like unsupervised learning, vision, planning, and statistical model selection. They will even be able to help us make decisions when the data is too complex for us to understand on our own.

And looking further ahead into the future, this could represent an important step in advanced computer bootstrapping — the ability for an artificially intelligent system to not just teach itself, but to re-write and improve upon itself. It could be seen as an important stage in the development of a recursively improving AI — a system that can continually become better at optimizing itself, potentially leading to an exponential increase in intelligence.

Image: The computer (yellow) carries out a new task (red); the program adds its prior training (green), makes predictions, and completes the task. The end result: a smarter machine. Credit: DARPA.