Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!rutgers!elbereth.rutgers.edu!harnad From: harnad@elbereth.rutgers.edu (Stevan Harnad) Newsgroups: comp.ai Subject: Re: Experience (was Re: Question on Chinese Room Argument) Summary: to come Keywords: The Robot Reply vs. the Robotic Functionalist Reply Message-ID: Date: 6 Mar 89 05:08:09 GMT References: <1989Mar4.152943.10902@cs.rochester.edu> Organization: Rutgers Univ., New Brunswick, N.J. Lines: 87 THE ROBOT REPLY VS. THE ROBOTIC FUNCTIONALIST REPLY yamauchi@cs.rochester.edu (Brian Yamauchi) of U of Rochester, CS Dept, Rochester, NY wrote: " The difference between a seeing man and a blind man is (surprise) that " the seeing man has functional visual sensors and the blind man does " not. Thus the seeing man can associate the linguistic term "blue" " with his sensory experience of viewing light in the "blue" frequency " range. There are many kinds of blindness, of which the lack of a retina (which is in any case not just a transducer) is only one. There are multiple analogs of the retina higher and higher in the brain. Their functions are not yet understood, but they seem to be doing feature extraction and transformation, all on projections that are isomorphic with the retinal surface. Blindness can arise from loss or deconnection at a variety of levels of brain function, including levels BEFORE and AFTER the last known retinotopic map. Nor are the nonretinotopic regions likely to be symbolic either. What's going on there is not yet understood. But it's certainly not the marriage of a symbol-cruncher to transducers. The nature of the functional transition from so-called "primary, sensory-projection" cortex to so-called "secondary, multimodal and supramodal association" cortex is simply not known by anyone at this time. But the progression seems to be from sensory input representation to multimodal activity to motor output patterns. No one has yet accused any part of the cortex of symbol-crunching (not even the "language" areas, Wernicke's area and Broca's area, which are receptive and productive, respectively). As to "associating" the symbol "blue" with the pertinent sensory category: It may seem easy in the case of color names (actually, it isn't), but when you move on to other concrete sensory categories, and eventually to abstract ones, the story becomes quite complicated. The problem of "connecting" the right symbol to the right input is the categorization problem, which in turn is closely related to what I've called "the symbol grounding problem." I have edited a whole book on this ("Categorical Perception: The Groundwork of Cognition." Cambridge University Press 1987) whose upshot is that it's not just a matter of hooking up a symbol cruncher to sensors and effectors. " The implication for AI seems to me to be that if we want systems which " can have "experiences" similar to our own, we need to equip them with " sensors which can perceive the physical world (or alternately, place " them in a very realistic simulated environment). You can give a symbol-cruncher sensors, but to make it PERCEIVE is a slightly taller order... " One could argue that the nature of human experience is also dependent " on the fact that human sensory inputs are processed as distributed " analog activations over a neural network (the brain). Personally, I " am undecided about whether this is a critical point. One safe way to remain undecided is not to worry about the problem of how to map symbols onto the world at all, simply assuming that it's easy, and that it will somehow meet a top-down symbol-cruncher half-way. Reasons and evidence are accumulating, however, to show what a simplistic pipe-dream that is, and how the very existence and self-sufficiency of an autonomous symbol-crunching module or level in the brain [or any other TTT-capable SYSTEM (sic)] may have to be rethought. One way of conceptualizing this is to contrast the standard "robot" reply to Searle ("You will have to connect sensors and effectors to your symbol-cruncher" -- which is easily parried by Searle, because it's irrelevant to his point) with my own "robotic functionalist" reply ("To pass the LTT you must be able to pass the TTT, and to pass the TTT you will have to draw upon nonsymbolic functions [e.g., transduction, analog transformations, A/D] which are immune to the Chinese Room Argument"). A functionally autonomous symbolic level may not exist in TTT-capable devices, much less be sufficient to give rise to LTT-capable performance. To put it another way, bottom-up may be the only route to the TTT, and the only way to arrive at a grounded symbol system. Ref: Harnad (1989) Minds, Machines and Searle. Journal of Experimental and Theoretical Artificial Intelligence 1: 5 - 25. -- Stevan Harnad INTERNET: harnad@confidence.princeton.edu harnad@princeton.edu srh@flash.bellcore.com harnad@elbereth.rutgers.edu harnad@princeton.uucp BITNET: harnad@pucc.bitnet CSNET: harnad%princeton.edu@relay.cs.net (609)-921-7771