| Humans are remarkably good at learning from instruction, even when that instruction is far from perfect. And while imperfect, natural human instruction often gets the job done. To get computers and robots to do what we want currently takes resource intensive programming and debugging by expert engineers. This is nothing like instruction. We seek to develop an "electronic student" that can learn from natural human instruction close to the way humans do. This technology will bring about a new generation of "field programmable" computers and robots, capable of being taught new concepts through natural human instruction. |
| To make human instructable computers a reality requires a number of advances over existing technology. The DARPA Bootstrapped Learning Program brings together researchers in machine learning, planning and control, and systems in order to develop this technology. |
| Under the rubric of noticing patterns and information that are not the explicit subject of the lesson being taught, we are developing methods for segmenting instruction dialog into thematic components based on teaching method and content, and noticing patterns in instruction interaction and the environment that are not direct subjects of teaching but provide critical foundations to the concepts being taught. |
People:
|
![]() |