EFTA02566518.pdf

From: S.M. Kosslyn Sent: Friday, February .... To: Joscha Bach Cc: Jeffrey Epstein Subject: Re: Today's discussion Hi J.. Some responses and reflections below... JJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJJ Stephen M. Kosslyn Director, Center for Advanced Study in the Behavioral Sciences Stanford University 75 Alta Road Stanford, CA 94305 Voice: On 14 Feb 2013, at 3:42 PM, Joscha Bach wrote: > Dear Stephen, > thank you for your so far unending patience in that discussion. --> > =ttp://www.xkcd.com/386/ That's a terrific website! >» I agree, and yet only a subset of animals can use symbols for =ommunication. » communication and intelligence are not the same thing (think about » =ees etc) >» Of these, only a small subset can make use of negation, conjunctions =nd disjunctions in symbolic communication (for instance, Irene =epperberg's famous grey parrots). It appears that only humans can learn =ich grammatical language, and I suspect that this is the primary =nabler of our superior problem solving capabilities. » I disagree. Einstein claimed that his greatest discoveries came from =ental imagery, and he later converted those thoughts to verbal =xpressions only with great difficulty. I think Al has vastly =nderestimated the role of "mental simulation/emulation" in thinking and =easoning. > Nonverbal thinking is primary, and is poorly understood, and it has > =een neglected by what we now often call "classic Al". (But if we look > =t the original ideas, that was probably not intentional. Logic based > =ystems were low hanging fruit. Once you have a paradigm and a > =ommunity, you end up with a methodology that is bound to stay, > =nfortunately.) EFTA_R1_01731748 EFTA02566518 That might be true.. but I think LISP lent itself to a certain way of =hinking.... > But nonverbal thinking is something that I suspect is quite similarly =owerful in other primates. I think we are much better at this than other primates; our conceptual =tructures are more powerful, and they in turn drive more powerful =ental simulations > I think that the most interesting difference between chimps and humans =s how we can use grammatical language to "program" and organize our =hought processes, and how we can use it to suggest, transmit, create =nd manipulate new categories. Perhaps. I'm not convinced. > I am on your side insofar as I think that the important research needs =o be done in mental imagery (or more accurately: on mental =epresentations and operations that facilitate mental imagery, among =ther things). But I think that human intelligence is shaped by the =dditions of grammars, which happen to be relatively easy to implement =hen you look at them in isolation. Only grammar on its own cannot do =hat Einstein did. Grammar is no doubt important, but I'm just not sure that it's at the =oot of what's most interesting about human intelligence. >» »» --thm)J]al]]1] So.. what would be wrong with building a machine =hat could do well on IQ tests? >» Nothing is wrong with building a machine that excels at playing =hess or cooking coffee or scoring that the Raven test. >> » The Raven would be a bad idea -- way too easy. The WAIS has some 11 » =ubtests, which cover a wide range of underlying abilities (and are » much =ore challenging) > Lets look at them (I have to admit that I am no expert on this, and it =s quite some time ago that I looked at IQ testing): > - The processing speed tests are probably trivial for computers If memory serves, none of the tests are about processing speed per se -- =hey are timed, but the issue is not simple processing speed, its =acility with certain kinds of reasoning > - The working memory tests are likewise rather simple engineering > =roblems Again, none of the tests specifically assess WM, although several tap =nto it. > - Perceptual reasoning is somewhat similar to the Raven (maybe I > =nderestimate them?) There are a set of perceptual reasoning tests, only some of which are at =11 like Raven > - Verbal comprehension: • - similarities and vocabulary tests are classical Al and =omputational linguistics • - information is close to IBM's Watson (recognition and =nference) 2 EFTA_R1_01731749 EFTA02566519 SO.. what you seem to be saying is that it would be simple to program a =omputer to do well on IQ tests. I would love to see this! > The only thing that looks interesting to me in the WAIS is the =omprehension test, because I don't see a straightforward approach to =heat on them with narrow Al. I would like to expand exactly this =omain: making sense of the world. Yes > We don't have advanced problem solving ("these are the rules for =hess. how would you try to beat a beginner level player, a medium =layer, a top player most quickly?"). We don't have constructive =bilities. We don't have verbal creativity etc. I think most of the above is in fact implicit in some of the tests. =emember that factor analysis reveals a very rich structure of human =ntelligence, with 60+ specific identifiable abilities that feed into =t. > Please tell me if my take on the WAIS is wrong! I think you might enjoy actually taking it. (My wife, when she was in =raining, used me as a guinea pig for testing -- and I found taking the =est really interesting... and was surprised by what I found trivially =asy and what I found more challenging) » How do you know for sure what the "basics" are? > While the literal understanding of the Turing Test leads nowhere (or, > =ell, to the Loebner prize), I think that he had the right idea. > =ntelligence is reflected in the ability to participate in meaningful > =iscourse, It may be reflected, but such discourse is not a necessary consequence =f intelligence. A deaf mute could still be very intelligent. > which includes interpreting and creatively structuring the world. Many =f the things that the WAIS measures, like recognizing and categorizing =hapes, are prerequisites for that. Others might be acquired tastes that =merge on more basic functionality, like mental arithmetic. But a =oolbox is not an architecture. A collection of tubes, tires, pedals and =pokes is not a bicycle. Good distinction. The IQ tests require a suite of skills and abilities, =hich could in principle arise from numerous architectures.. > Some of the basics stem directly from the requirements of producing =dequate representations of perceptual and abstracted content (hybrid =ierarchical representations that can do associations, compositional =tuff, grammatical systematicity, learning and categorization, =nheritance etc.). Others come from the needs to get the processes of =ottom-up/top-down perception, reflection, memory retrieval, inference, =nalogy building etc. to work. And some have to do with the requirements =f translating between Mentalese (in Pinker's sense, not in Fodor's) and =atural language. I don't disagree with any of the above > I may delude myself in thinking that I know what the basics are. In =act, it is extremely likely that I do (every computer science problem =eems to be misconceptualized until it has been properly implemented). You must be familiar with what the classic Al guys (e.g., Herb Simon) =alled "the representation problem" 3 EFTA_R1_01731750 EFTA02566520 > But I would start with mental representation, perceptual processing =nd motivational relevance, and then go for language, while revisiting =hose areas that turn out to fall short. I would stop before language, but this may reflect a deep prejudice on =y part. I think that much of logic comes out of perceptual experience =ith contingencies in the world » Forget about the Raven; it's a non-verbal test of fluid intelligence » =which in fact turns out to have, by accident not design, two » different =ypes of items -- solved by spatial vs. analytic » strategies). The Raven =oes not even begin to characterize all of » what is captured by the WAIS > It might well be that I totally misunderestimate the WAIS "requirements; I will look at them. Better yet: Have somebody actually give it to you. The actual WAIS =annot be taken on a computer or the like; it needs a trained person to =dminister it >» I agree. But I am not convinced that proper emotions are absolutely =ecessary for Intelligence (motivation might suffice to drive some kind =f non-emotional, serene Buddha intelligence). I am nevertheless =nterested in understanding and modeling them. » I think Antonio Damasio and his successors have made a very good case » =hat emotion plays a key role in reasoning. (Not just motivation, » actual =motion.) > Most of what I would say against that has been better expressed by =aron Sloman. For instance, if my computer is prone to swapping memory =ontent to hard drive and back, and I kill the part of the OS that =oordinates the swapping, my computer is likely to malfunction. But from =his I cannot infer that computers cannot work without swapping. I believe that Richard Gregory had a version of that argument well =efore Sloman > Damasio's argument does not convince me because he does not elucidate = functional role that would emotion an absolute requirement for an =rtificial mind. I think emotion serves to prioritize goals, which seems kind of =mportant > Lesion studies amount to shutting down parts of an operating system =hat has been designed to cope with very specific requirements. I =elieve that in humans, emotions structure social interaction, support =ommunication, prime memory and cognitive processing, and most =mportantly, allocate the scarce resources of our mind according to the =urrent situation. None of this is necessary if I remove the resource =onstraints. Lesion data were the beginning, but there have been many sorts of =tudies now of the role of emotion in reasoning. The "social =euroscience" folks have done a lot of this stuff. I'm convinced. > But again, perhaps there are better arguments now than in the original =omatic marker hypothesis? I think so... I recall Kevin Ochsner years ago telling me how lousy the =ard test was, and that there were much better ways to get at the same =oint (but I no longer recall what those studies were.. ) > Please do not misunderstand me; I am making a merely philosophical =oint here, with respect to the basic requirements for Intelligence. I =hink that emotion is highly interesting, that Damasio is quite correct =ith respect to 4 EFTA_R1_01731751 EFTA02566521 what emotion does, and that it makes a lot of sense (and =s fun) to equip Als with emotion, mood, affect and emotional =ispositions. But strictly necessary? No. I disagree; I think emotion is crucial for rapid interrupts and setting =riorities (yes, motivation is also involved, but generally has a longer =ime horizon) » Are you involved in BICA? That seems like a natural community for =ou! > The way I understand it, there are at least four very similar groups > =ow: cognitive modeling (that is where John Anderson goes), AGI > (started =y Ben Goertzel as an attempt to revive the original AI) I don't know about this > , BICA (a remnant from the failed DARPA proposal of the same name, and > =ater picked up by Alexei Samsonovich as an alternative to AGI, I > =uspect because he does not get along with Ben) Yes, the history is correct -- I was part of one of the original teams. > , and Cognitive Systems (Pat Langley et al.). I basically like them > =11, and think that they should join forces, while simultaneously > =aising the bars against narrow Al and science fiction. Many members > of =he audience already belong to two or even three of the groups. > Alas, =olitics, mutual accusations of scruffiness and stuffiness, and > so on... Weird. I had no idea that AGI or "Cognitive Systems" existed... Another =easure of how out of touch I've become... > Personally, I have not been to one of the BICA conferences (only a =ouple planning workshops), and I am on their roster of reviewers. They have a journal now; it might be worth a glance.. Be well! s. > Cheers, > Joscha <?xml version=.0" encoding=TF-8"?> <!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd"> <plist version=.0"> <dict> <key>conversation-id</key> <integer>245038</integer> <key>date-last-viewed</key> <integer>0</integer> <key>date-received</key> 5 EFTA_R1_01731752 EFTA02566522  <Integer>1360902239</integer> <key>flags</key> <integer>8623750145</integer> <key>gmail-label-ids</key> <array> <integer>6</integer> <integer>2</integer> </array> <key>remote-id</key> <string>276496</string> </dict> </plist> 6 EFTA_R1_01731753 EFTA02566523