DARPA leads new AI research
Computer scientists have long sought to develop computers that can match the subject expertise that humans acquire during a career or a lifetime. Despite intensive work with expert systems and other forms of artificial intelligence, researchers have discovered that building a computer that can learn like a person is more difficult that they expected.
Now, with a Defense Advanced Research Projects Agency (DARPA) program called Bootstrapped Learning, the agency wants to generate renewed interest in achieving that objective. SRI International recently won a $10 million contract to lead the first 15-month development phase of the program.
Introducing the initiative last year, DARPA Program Manager Dan Oblinger called Bootstrapped Learning a completely new approach to machine learning.
The goal is to allow humans to teach computers in a way thats natural for humans
and then for computers to be able to assimilate the knowledge they gain, said Roger Mailler, a senior computer scientist and deputy principal integrator at SRI. Its the reverse of the way other programs have approached machine learning, he said.
A primary advantage of the bootstrap approach is that computer systems could be programmed by people who have critical knowledge in specific areas but who arent necessarily computer experts, he said.
The operation of unmanned aerial vehicles (UAVs) is an example of an application for the technology, Mailler said. To carry out missions, UAVs must be programmed with new behaviors to correspond to various flight characteristics. Each time mission requirements change, the UAVs must be shipped to their bases to be reprogrammed and undergo months of testing before they can be redeployed. When requirements change again, the entire process must be repeated.
This results in operations always being behind the curve, Mailler said. With bootstrapped learning, however, people in the field who are subject matter experts in UAV flight control would teach the UAV what it needs to know, even though they themselves know nothing about computer programming.
In other settings, bootstrapped learning technology might assist people trying to gain access to information stored in complex computer systems. For example, a person must have extensive knowledge of a particular database system to be able to extract relevant data. With bootstrapped learning, people with no database knowledge could describe what they need, and the computer would know how to retrieve the relevant information.
The objective of the SRI-led first phase of DARPAs Bootstrapped Learning program is to develop a learning system called Phased Learning through Analyzing, Teaching and Observation (PLATO). The result will be a domain-independent electronic student that can learn from human instructors, understand the implications of that instruction in a particular context and be able to refine that learning over time, as necessary.
The machine will be capable of learning how the instructor teaches and to understand what task is being conveyed to it, Mailler said.
Project leaders hope that such systems will be easier and less expensive to maintain than current AI systems and be faster to update and deploy.
The second phase of the Bootstrapped Learning program, for which contracts have not been awarded, will be to develop a simulated person that can teach the electronic student.
A field-deployable system based on bootstrapped learning developments could be introduced in the next 10 years, Mailler said. But whether that happens will depend on how the AI community and others respond to the challenge. Achieving that goal will depend on contributions from many sources, he said.
Were developing the basic electronic student capability to allow the public to work with it, and to invigorate the research community to work on this, Mailler said. We need them to recognize that this is a first-class problem that needs to be addressed.