DARPA wants to build a learning computer

Hello, Professor Falken.

DARPA is a division of the U.S. Department of Defense tasked with developing new technology for use by the military. Their most famous project was called ARPAnet, which eventually became the Internet. They define the term “scary smart,” and while some of their technology trickles down to the civilian sector, most of it is designed to enable the U.S. Military, which sometimes involve helping people and building things, and at other times, killing people and breaking things.

Dr. Robert Kohout, an expert in artificial intelligence, is developing a program called “PAL” for DARPA’s Information Processing Techniques Office.

The PAL project has five main goals:

  1. Enable machines to learn and improve their basic functionality through the accumulation of experience (and not through being explicitly programmed).
  2. Can represent purpose/goals, system structure, and behavior, in order to allow a computational system to reflect on its own capabilities and performance.
  3. Allow the software to be instructed and guided using natural human-oriented communications (e.g., natural language, pictures, gestures).
  4. Have the ability of the software to use visual and auditory sensors to understand the user’s situation (who is in the meeting, who is speaking, etc).
  5. Are integrated and result in fully functioning systems.

In other words, they want to build a learning computer that can understand English. And they want to hook it up to military computers.  I’ve seen this somewhere before

It seems like science fiction, but the PAL program literally list its mission as enabling systems that “can reason, learn form experience, be told what to do, explain what they are doing, reflect on their experience, and respond robustly to surprise.”

And, thankfully, DARPA has provided a video illustrating the concept:

Advertisement





6 Responses to DARPA wants to build a learning computer

  1. there are major issues with a computer that can learn. in terminator Skynet started as a learning computer integrated into the World Wide Net but the computer could learn. A computer will fulfill its directives to the letter but if it can learn and it learns the the only way to fulfill its primary directive is to destroy mankind and its integrated into all our weapon systems than it will launch a preemptive strike on key targets like Washington DC and NORAD. Computers can become "fully aware" and i believe if that happens than certain depictions of events could become a reality… With disastrous consequences.