From weng@cse.msu.edu Fri Mar 3 16:27:20 2000 From: Dr John J Weng Date: Fri, 3 Mar 2000 16:27:14 -0500 (EST) To: wdl@cse.msu.edu Subject: WDL Dis. #12 ======== WDL Dis. #12 (Bcc: all WDL participants) ====== Stan, > I believe that some amount of prior knowledge must be built in either by > a designer or by evolution in order for any learning to occur. Your comments touched an issue that is very fundamental but its common answer has unfortunately mislead many AI researchers like me for too long. What is innate in a higher organism? Prior knowledge or something else? I guess that Kim Plunkett can give a more detailed account when he gives his review talk at WDL. Here I recall from the book ``Rethinking Innateness'' that he co-authored with other five researchers. They outlined three levels of the nervous system in a higher organism: (iii) Representation level (e.g., weights on neural connections) (ii) Architecture level (e.g., neuron types) (i) Timing level (the time schedule of neural growth during development) They do not object to the view that a lot in (i) and (ii) are innate, but they hold a view that innateness in (iii) is very rare in higher organisms. I am a fan of neuroscience. As far as I can see, (iii) above is too rough and it is a misnomer. In a book that I am writing, I broke it into three to become levels (3) to (5): (5) Knowledge level (e.g., object permanence*, linear ordering, symbolic skills, thinking skills, general understanding of the world around us, learned part of emotions, and rich consciousness) (4) Inborn behavior level (e.g., rooting, sucking, breathing, pain avoidance and some primitive emotions in neonates. In neurons, they are related to synapses at the birth time.) (3) Representation level (e.g., how neurons grow based on sensory stimuli) (2) Architecture level (corresponding to anatomy of an organism. E.g., a cortex area is prepared for eyes, if everything is developed normally) (1) Timing level (the time schedule of neural growth of each area of the nervous system during development) Studies in neuroscience seem to show that all of the above 5 levels are experience-dependent. In fact, experience can shape all these levels to a very great extent. But it seems that our gene has designed a lot for (1) through (4). (5) is made possible by (1) through (4) plus experience; but (5) is NOT wired in. Thus, (1) through (4) seem what a programmer for a developmental algorithm may want to design --- but not rigidly --- experience dependent. >From a design point of view, (5) is the MOST difficult one. It seems more systematic and EASIER if we look into (1) through (4). E.g., we know that the subjects of vision and intelligent robots are very difficult, but they should not be that difficult if we look into (1) through (4) instead and make our robots develop automatically. The same paradigm seems applicable to psychology. Psychology will become a much beautiful science if the field shifts its current emphasis on (5) to (1) - (4) instead. (5) is very ad hoc by nature. A lot questions of long debates in psychology could be answered systematically if psychology searches answers through (1) - (4). Any comments from participants who like to complement above or dispute again it? Stan raised a great issue ... John --------------------------------------------------------- * What is ``object permanence?'' We human adults believe that objects are external to us; they do not change their identity when they are moved to another location and they continue to exist when they are moved out of our sight. Piaget called this understanding object permanence. He reported that infants lack the sense of object permanence till about 10 months of age. I guess that Esther Thelen and other developmental psychologists will explain more fascinating studies about of child mental development at our WDL.