Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!csd4.milw.wisc.edu!lll-winken!uunet!mcvax!hp4nl!htsa!fransvo From: fransvo@htsa.uucp (Frans van Otten) Newsgroups: comp.ai Subject: Re: Where might CR understanding come from (if it exists) Summary: An explanation on the different meanings of "understanding" Message-ID: <811@htsa.uucp> Date: 23 Mar 89 11:43:44 GMT References: <28867@sri-unix.SRI.COM> Organization: AHA-TMF (Technical Institute), Amsterdam, Netherlands Lines: 153 Michael Ellis writes: >Wayne A. Throop writes: > >>The systems reply is that Searle-plus-the-rules *does* understand. > >That requires a leap of faith I'm willing to make for other >humans, animals (and martians should they arrive), but not >for an artifact whose design fails to include whatever >relevant causes there may be to consciousness itself. The word "understanding" has three very different meanings. I think the main problem in this entire discussion is that most people keep mixing those meanings. *** First interpretation of "understanding" I can say "I understand". This statement is based on some feeling; it is subjective. This might be represented/implemented by a flag which is set to "true" or "false", or by a variable which contains a value within a (continuous) range, where there is some critical value when I get the feeling that I do understand. *** Second interpretation of "understanding" I can also say "You understand". This refers to some other entity then me myself. This is the "understanding" which is "measured" by the Turing Test (whatever variant you are using). Then first of all, there is the "other minds problem". Does the entity I am referring to have a mind ? I don't know. But I do say that it understands. So I might conclude: External behaviour resulting in me stating "you understand" does NOT require that the entity has a mind. And there was the argument "when you know the used algorithm, you usually don't say that it understands". I disagree on this one. Let's take a calculator. I can say "it understands how to add two numbers". Let's take an email-program. When I mail to someone on my local computer, the program "understands" not to call some other system for that message. When I mail to some unknown system, it does understand that the message must be sent to some backbone. My point is that in the process of thinking about things like "understanding", "consciousness", "intelligence" etc. many people lose their normal interpretations of those words and start to require undefined things, just because "human understanding is so very special, how could some simple device like a computer ever be able to have it ?". *** Third interpretation of "understanding" Neither of the above described interpretations of "understanding" mentions the actual process involved in understanding. There is no hint as to how understanding might work, what it might be. The interpretation of "understanding" concerns this process or state, the actual implementation of understanding a concept. The Chinese Room argument mixes the three described meanings of "understanding" in such a way that nobody knows what is true and what is false: 1. We have a set of rules [a computer program] defining which Chinese characters should be presented in response to incoming Chinese characters. It is assumed that when someone strictly follows those rules, the other person(s) involved in this conversation could be fooled into believing they actually are having a correspondence with a human being (or some other entity) understanding what is written. This concerns the second interpretation of understanding I described: "you understand": the Turing Test is passed. 2. Then the argument continues: "the person interpreting the rules doesn't understand Chinese". This concerns the first interpretation of understanding (the subjective one). 3. Then finally, the conclusion is made: "as this person, who is doing everything a computer would do, does not understand Chinese, then a computer can't understand Chinese either". This refers to the third interpretation of understanding, the objective one, the kind AI-researchers are (or should be) interested in. Objections against this argument have been: 1. The set of rules would be far too big, too complex to be executed by a human in a reasonable amount of time, etc., the argument relies on intuitions which are misguided thus. Of course, the set of rules would be highly complex. But the Chinese Room argument was a thought experiment, so that is no problem (but it helps the confusion). Would it have been possible for Einstein to travel on a lightbeam ? 2. The Turing Test tests on bevioural characteristics of the system, not on the internal (cognitive) functions. Very true, but that was exactly why the Turing Test was used in the argument. The idea was "make it seem to understand, this can be measured by the Turing Test, then see inside to find out whether or not it does understand". 3. The person executing the rules is not able to feel the understanding of the system. That is true, too. This is one of the flaws of the argument. But it doesn't touch the heart of it. We are not looking for a system that has a feeling that it does understand. We are trying to find a system that (objectively) understands. My objection against the Chinese Room argument is that the different interpretations of the word "understanding" are mixed in a way that is not tolerable. First of all, behavioural characteristics tests as the Turing Test cannot produce any evidence. The best they can do is "it seems to understand" or "it doesn't seem to understand". Drew Mcdermott was very clear on this is his paper he posted on the net. Secondly, as many posters wrote, you can't base a conclusion about an entire system on observations (let alone feelings) of a single part of it. This has been shown in many ways. The best way was probably "let someone execute the rules of physics of his own body; although the person understands things, he wouldn't understand those things from calculating his own physics". Note: The question has been raised, if the person would understand the rules, would he then understand Chinese ? I think Stevan Harnad is right on saying "no"; the symbol grounding problem does exist ! When you know that you should respond with a certain Chinese character when you see another certain one, this doesn't make you understand Chinese ! My theory on this is, that understanding requires representation of the external symbols (be it Chinese characters, vocal representations of words, red traffic lights, or whatever) in internal symbols. You can base actions on your internal symbols, but not on external symbols. Then the "denotation/connotation" argument of Roel Wieringa shows up. How do I translate external symbols to internal symbols ? The process of reaching a conclusion based on an internal state (I call this process "intelligence") is independent of the external symbols and their meanings. But the rules that are used to reach a conclusion are based on the socially accepted meanings of the external symbols. The same holds for the translation from external to internal symbols. This is what Karl Kluge meant when he wrote about communicating his desire for a bowl of icecream to his appartmentmate. Final note: If you ever manage to make a system that does understand, pack it in a green humanoid container, put this in a rocket, write in big letters "From Mars" on it, and launch this rocket to Michael Ellis. -- Frans van Otten Algemene Hogeschool Amsterdam Technische en Maritieme Faculteit fransvo@htsa.uucp