Hacker News Reader

Comment by deepnet | Hacker News Reader

Worth noting this Socratic review of Sir Penrose's theory of quantum minds is by Dr Hans Moravec, an important roboticist and AI theorist, widely known for Moravec's Paradox :

"it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility."[1]

which began the shift in paradigms exemplified by the behaviour based & embodied robotics movements.

  [1] Dr Hans Moravec, snipped from http://www.eugenewei.com/blog/2014/10/13/moravecs-paradox-and-self-driving-cars

View on HN

Playing checkers seems like a well-defined task to accomplish. A computer that has "the skills of a one-year-old when it comes to perception and mobility." is not defined at all. Mobility? Perceiving what?

Im reminded of the multimillion dollar AI-based image recognition system the military developed, in the 80s or 90s to analyze images looking for tanks. When the system was tested on photos, that were not from their pool of stock test photos used to train the systems, it got almost 100% wrong. Why? Because they stock photos had well-composed shots on a sunny day with some clouds (usually the spots tanks are hidden in are overcast or very cloudy). The "real" shots were all cloudy (as expected) and the system had actually been trained to recognize discrete clouds rather than Tanks.

Perception is about distinguishing, imagining physical properties (this can be disassembled or just casually appears composite due to perspective), judging and guessing at unseen properties. I figure an pool ball has a curved 3d spherical shape, because I've seen one before and felt it. At the same time I imagine that there is no "top" or differing feature anywhere on the ball, based on the appearance of what portion I can see (it looks clear and flat white? i expect the rest to be so).

To clarify, Moravec was speaking of robots, e.g. embodied computation.

Moravec's paradox, though it now seems obvious, was the opposite of the prevailing view.

It implies that if 'being in the world' (mobility, perception & grasping) are the hard tasks then this is where AI will emerge not in simulation, planning or board games.

My memory is failing me, but I think I heard a talk once where a (specific?) robot motion-planning problem is PSPACE-hard, aka in a class that contains the problems in NP. Maybe that's another way to articulate Moravec's paradox.

Reply on news.ycombinator.com