Full-text

URL: http://www-eksl.cs.umass.edu/papers/cohenIAS00b.ps

Cached: PDF-27K    PS-81K    PS.gz-26K

SAVE AS an easy-to-recall long filename:

Filename format: author--year--title   PDF-27K    PS-26K    :: About GZip'd PS

Filename format: author--year--title--journal|proceedings|...--pages   PDF-27K    PS-26K

Related links

CiteSeer: http://citeseer.ist.psu.edu/261054.html

Web search: Google Web Search   ::   Google Scholar

Within this site: References (15)

Paper at a Glance

Learning Concepts by Interaction
Paul R. Cohen
Department of Computer Science
University of Massachusetts, Amherst, MA 01003
cohen@cs.umass.edu

Abstract This paper presents a theory of how robots may learn concepts by interacting with their environment in an unsupervised way. First, categories of activities are learned, then abstractions over those categories result in concepts. The meanings of concepts are discussed. Robotic systems that learn categories of activities and concepts are presented.
Introduction If machines could acquire conceptual knowledge with the same facility as humans, then AI would be much better off. There's no denying the dream of a machine that knows roughly what we know, organized roughly as we organize it, with roughly the same values and motives as we have. It makes sense, then, to ask how this knowledge is acquired by humans and how might it be acquired by machines. I focus on the origins of conceptual knowledge, the earliest distinctions and classes, the first efforts to carve the world at its joints. One reason is just the desire to get to the bottom of, or in this case the beginning of, anything. Another is that the origin of concepts is hotly debated: Some people think that neonates are born with moderately sophisticated conceptual systems (e.g., 4,5,15), others dispute this and seek an empiricist, or non­nativist, account of development (e.g., 10,11). I think one has to take a minimalist stance and avoid innate knowledge in one's explanations of the acquisition of later knowledge. This reluctance comes, in part, from years of slogging in AI, where we must always provide a lot of knowledge for our systems to do relatively little with. I want to focus on the first concepts because unless I do, I will have to provide them by hand, which is difficult, and also makes me very uncertain about the explanatory power of what follows. The claim of this paper is that concepts can be learned without supervision by abstracting over ...

@techreport{cohen00learningConcepts,
  author={Paul R. Cohen},
  title={Learning Concepts by Interaction},
  year={2000},
  institution={Computer Science Department, University of Massachusetts at Amherst},
  note={Technical Report 00-52},
  url={http://www.isrl.uiuc.edu/~amag/langev/paper/cohen00learningConcepts.html}
}

HOME   ::  Technical Report List   ::   Technical Report	Comments to: junwang4 you-know-at gmail.com	Last update: 10/27/09