Ben: Google is not trying to do AGI, they're explicitly doing statistical machine vision... so what?
Quick response to you & Jim, mainly to throw out a hypothesis for investigation. The "so what" is this - while I take your point, Ben, I doubt that any AGI exercises, a la Poggio & Arel, are actually going to be any better - once you allow for the fact that AGI-tests are always on much more heavily controlled & selected/ "configured" materials. Do you have any tests that show any real AGI promise whatsoever? The thesis I wish to play with is that just as words reduce to letters, so do visual forms reduce to a basic alphabet of abstract forms (albeit not a precise, closed alphabet) - which would include such basic units as the blob, wedge, chunk, (irregular), and the circle, triangle, square, (regular) and straight and zig-zag or wavy lines. And a Google or more AGI exercise of recognition may be able to, with some mild but unimpressive success, identify similar basic forms. But just as being able to recognize words has v. little to do with being able to understand the concepts they label, so does being able to recognize abstract forms have v. little to do with being able to recognize the real bodies/. body forms in visual scenes, of which those basic forms are components (as Google Images demonstrates). Being able to recognize a circle, to put it v. simply, is a world away from being able to recognize whether it's a disk, tyre, ring, open mouth, sun, moon, plate, roundabout, geometric circle, etc.- or whether a straight line is a road, rod, tube, pillar, cable, finger, etc. or a wavy line a wave, snake, or rumpled carpet or water bed. Visual body recognition requires a robot/agent with the capacity to deal imaginatively with whole complex body forms as well as component, abstract forms, (such as those listed), and register and understand their v. different principles of body movement. That would include understanding the v. different principles of movement of disks, tyres, suns, plates, mouths etc. and how they got into their place on a screen/scene and are likely to proceed from that place. Visual body recognition requires a world of both imaginative and lived experience of the very different bodies in this world. Plus the capacity to *imaginatively* (not geometrically) reconstruct and play with that world of bodies. Plus the capacity, as with language understanding, to embody those bodies with one's own body, and not just play with them on a screen. (Way, way beyond your "Manhattan project"). P.S. I was looking for a phrase as I wrote this & I may have got it - it's not enough to know about the physics and biomechanics of movement, in order to understand bodies - you have to know about the ART OF MOVEMENT. Human and animal movement are indeed an art - both in terms of production and the final forms of movement. Knowing about biomechanics will not help you understand or predict the infinitely open-ended range of dances that humans keep producing. Understanding and creating those dances is literally, technically, an art/art form - just as drawing dancers is. ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
