2010.01.15 - Benjamin Kuipers - How can a robot learn knowledge?

    • Goal of AI: mathematical model of mind
    • Successful AI: foundations of knowledge must be learned
    • Syntactic AI has problems - Chinese Room by John Searle
      • But computation by nature is syntactic
    • Constructivist AI: constructs its own internal knowledge
      • We're OK to assemble our own delusion as long as it corresponds to the real world, otherwise we die.
    • To accomplish constructivist AI we do bootstrap learning:
      • Body space
      • Local spatial model
      • Objects and actions
      • Goals, beliefs and plans
      • Other "agents"
    • We can use the idea of distinctive states to help here:
      • Get in the general vicinity of the goal and then refine
    • Example of constructivist learning:
      • Estimate similarities between random sensory input
      • Learn organization of sensory information
        • First an ego-centric view
        • Then a world-view
      • Learn objects
        • These are discrepancies (mutable things) in the world view
      • Learn actions
    • A true AI would be able to do the above example with all of the above rules running in parallel.