Ben,
Good Post I my mind the ability to map each of N things into a model of a space is a very valuable thing. It lets us represent all of the N^2 spatial relationships between those N things based on just N mappings. This is something we all know, but it is one of the many wonderful efficiencies of mathematics we often don't stop to appreciate. Ed Porter -----Original Message----- From: Benjamin Goertzel [mailto:[EMAIL PROTECTED] Sent: Sunday, October 21, 2007 8:49 AM To: agi@v2.listbox.com Subject: Re: [agi] Human memory and number of synapses.. P.S. Some semi-organized responses to points raised in this thread... 1) About spatial maps... It seems to be the case that the brain uses spatial maps a lot, which abstract considerably from the "territory" they represent Similarly in Novamente we have a spatial map data structure which has an adjustable level of detail -- but for most purposes a high level of detail isn't needed. (Most obviously, we use spatial maps right now for navigation within virtual worlds; but will also use something similar for object recognition/classification....) In Novamente a spatial map is basically a spatial index into the overall AtomTable that stores knowledge as nodes and links. So we have a table of spatial maps, each spatial map living in its own relative coordinate system. 2) About the level of detail remembered about visual scenes Furthermore, it seems to be the case that the brain stores a lot of detail about some things that it sees -- and much less about others. For instance, it's famous that when observing a visual scene, a person can generally remember only around 7 visual facts about it. Trained observers can of course overcome this limitation. But there are loads of psychological studies validating how little we really remember of scenes we've seen. Check out the book "Is the Visual World a Grand Illusion?" which is an informative little edited volume. 3) About the constructive nature of memory In very many cases, what we feel like we're remembering is actually stuff we're reconstructing. This is validated in so many domains and ways via psychological experiment, I don't have time to dig up the references now. It's textbook cognitive psych. This applies in vision as well. We think we're remembering a detailed visual scene, but really we're remembering something sketchy and filling in the details. 4) On the value of visual imagination for cognition So, here's one question. Let's say we're imagining something visually in the course of thinking about it. And let's say that this visual imagination consists of -- some dynamics going on at the level of an underlying, abstracted representation (the form in which we really remember most visual scenes, for example) -- some dynamics going on at the level of "constructed details", invented on the fly based on this abstracted representation How much thinking has to do with the underlying representation, and how much with the constructed details? I would bet the vast majority has to do with the underlying representation, yet the constructed details may also play a role, say, in activation-spreading between associated things... 5) Vision is not central to cognition So, I do think that for any mind associated with a system that has to move around in a complex world, and that uses sight as a primary modality, the manipulation of spatial maps and associated images is going to be an important aspect of cognition. However, I really don't think it's central. I think you could make a humanlike mind that lacked vision and instead possessed some totally different kind of sensorium. I think the human brain architecture would still make sense. Similarly, an architecture like Novamente still makes perfect sense if one gets rid of vision and replaces it with something else. 6) The cortex does not "speak vision"; the cognitive cortex did not evolve from the visual cortex It's worth noting, finally, that it is OLFACTION not vision that feeds directly into the cortex. Visual stimuli, like that of the other 3 senses, must pass thru the thalamus to get translated into cortex-language (so to speak). A lot of preprocessing in the LGN and other places is required to translate the output of the optic nerves into something the cortex can grok. So it's not the case that the cortex is somehow naturally structured to reflect the structure of visual data. More so it's naturally structured to reflect the structure of olfactory data. Or so says neuroanatomy... Furthermore, cognitive cortex evolved from olfactory cortex not visual cortex. As Gary Lynch noted years ago, this is likely because olfaction involves a lot of wildly tangled combinatory neural connections, whereas vision (as Jeff Hawkins is fond of observing) involves a lot of hierarchical connections. I suspect that the wild intercombinatory aspect of olfactory cortex proved very useful, evolutionarily, for cognition ... more so than the more rigid hierarchy of visual cortex. 7) It's just not that simple... As realizing a mind within limited computational constraints is difficult and complex, it's very tempting to take a single aspect (logic! images! reinforcement learning!) and decide this aspect is the essence of intelligence, or human intelligence. I don't think so. As I said long ago (and others said longer ago), the key thing is to have an architecture in which different semi-specialized modules can play together and learn from each other and modulate each others' intrinsic inefficiencies and shortcomings. The human brain provides this (see Steven Mithen's book on the Prehistory of Mind for some ideas on this from an evolutionary perspective) and the Novamente design tries to provide this also... -- Ben Goertzel _____ This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/? <http://v2.listbox.com/member/?&; > & ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244&id_secret=56053709-a327cd
