tag:blogger.com,1999:blog-3722233.post411970884094691767..comments2024-03-28T18:17:00.135-05:00Comments on Computational Complexity: Starting the Year with TuringLance Fortnowhttp://www.blogger.com/profile/06752030912874378610noreply@blogger.comBlogger14125tag:blogger.com,1999:blog-3722233.post-26321870558187638552012-01-08T20:46:37.589-06:002012-01-08T20:46:37.589-06:00I was also at Minsky's talk. I thought it was...I was also at Minsky's talk. I thought it was embarassing, mainly because it was billed was a talk on how Turing had influenced everything Minsky and his students had done, and it seemed to me that it didn't really touch on that (as well as being rather rambling and unfocused). I did subsequently speak to someone who thought that Watson was significant because of how it could process natural language -- which is another question entirely, of course, but interesting and significant in itself.Carl Offnernoreply@blogger.comtag:blogger.com,1999:blog-3722233.post-46095823574099214762012-01-08T18:38:30.142-06:002012-01-08T18:38:30.142-06:00Anonymous said... If there was [a narrative founda...<i><b>Anonymous said...</b> If there was [a narrative foundation for human chess mastery], humans would learn chess by memorizing said narrative, instead we do massive data collection and (subconscious) classification."</i><br />--------<br /><br />Granting that massive data collection is necessary to chess mastery, it is striking that chess grandmasters <i>also</i> are notably skilled at annotating games, and conceiving stories about the ebb and flow of tension in the game. Bill Thurston's celebrated article <i>On proof and progress in mathematics</i> surveys similar themes.<br /><br />Perhaps what distinguishes (today's) computer programs like IBM's Watson, or even <a href="http://www.ncbi.nlm.nih.gov/pubmed?term=22206802%5Buid%5D" rel="nofollow">the computational capabilities of human savants</a>, from the capabilities of masters like Kasparov and Thurston, is that the latter have a facility for collecting data and refining narratives <i>simultaneously</i>.John Sidleshttps://www.blogger.com/profile/16286860374431298556noreply@blogger.comtag:blogger.com,1999:blog-3722233.post-32700094759342533632012-01-08T14:54:12.145-06:002012-01-08T14:54:12.145-06:00and perhaps this because we humans haven't (ye...<i>and perhaps this because we humans haven't (yet) constructed good narratives about how we construct narratives (chess or otherwise).</i><br /><br />What makes you believe that they exist? Fifty years ago this would have been a not bad working hypothesis, but the lack of a single one in the ensuing half-century of AI research strongly suggests that there is none. If there was one, humans would learn chess by memorizing said narrative, instead we do massive data collection and (subconscious) classification. <br /><br />We seem to be a collection of simple, one-of heuristics that the brain can deploy massively in parallel, including semi-exhaustive searches of choice branches of the solution space, most of which happens subconsciously.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-3722233.post-91678372829254205352012-01-08T14:45:57.238-06:002012-01-08T14:45:57.238-06:00Feel bemusement? some of those grandparents should...Feel bemusement? some of those grandparents should feel embarrassment for setting AI out on a wild chase of overpromising and underperforming research lines. <br /><br />Minsky's "advice" about how and how not to do AI research---which are unchanged since his early years---should have about the same weight as the statements of a XV century Chinese rocketeer lamenting the sorry state of today moon rockets because they lack the sparkles his come to expect from fireworks.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-3722233.post-15109751011281469272012-01-08T12:02:45.288-06:002012-01-08T12:02:45.288-06:00The great grandparents of AI — Turing, McCulloch, ...The great grandparents of AI — Turing, McCulloch, Pitts, Minsky, Papert, Ashby, McKay, Weiner, Chomsky, Schützenberger, Arbib — were at least literate about the classical dimensions of the enterprise. I can imagine how bemusing it would be for them to observe the same old hammers being used on the same old screws.Jon Awbreyhttp://mywikibiz.com/Directory:Jon_Awbreynoreply@blogger.comtag:blogger.com,1999:blog-3722233.post-23372933396748143852012-01-07T10:28:31.649-06:002012-01-07T10:28:31.649-06:00Minsky talked mostly about the sorry state of AI o...<i>Minsky talked mostly about the sorry state of AI over the past few decades including how the Watson people were working on the wrong problem. </i><br /><br />I'd say that doing the opposite of what Minsky says is a sign that AI is on the right track.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-3722233.post-37741772722810487102012-01-07T08:30:55.561-06:002012-01-07T08:30:55.561-06:00We can imagine a Watson-style AI that is expert in...We can imagine a Watson-style AI that is expert in playing chess: composed of a huge database of openings, a vast set of heuristic evaluation rules, a fast alpha-beta search engine, and an enormous lookup table of perfect end-game strategies. Of course there is no need to imagine such engines; they exist concretely and their code is in the public domain (Crafty, Fruit, Stockfish).<br /><br />Then we can imagine a Minsky-style AI that can play chess <i>almost</i> as well as the Watson-style AIs, and moreover can do what the Watson-style AIs cannot do: construct explanatory narratives about chess games. For these AIs, the narratives <i>are the code</i>. Of course there is no need to imagine such AIs; they exist concretely as chess grandmasters, and their narratives are largely in the public domain (Kasparov's <i>My Great Predecessors</i>).<br /><br />As is well known, machines aren't (yet) much good at constructing narratives-as-compact-codes (chess or otherwise), and perhaps this because we humans haven't (yet) constructed good narratives about <i>how</i> we construct narratives (chess or otherwise).<br /><br />In this latter regard — the process of constructing compact narratives about the process of compact narrative construction — I commend to readers of the Fortnow/GASARCH weblog Amelie Rorty's celebrated (in some circles) and much-reprinted article <i><a href="http://books.google.com/books?id=KOftYyztQXYC&pg=PA65" rel="nofollow">Spinoza on the Pathos of Idolatrous Love and the Hilarity of True Love</a></i>.<br /><br />To the extent that Rorty's narrative conceptions could be formalized and instantiated in machine code, considerable progress might be made toward post-Watson modes of self-programming AIs. <br /><br />In any event, the capability to read-and-understand essays like Rorty's amounts to a Turing test that no present-day machine (and perhaps not too many present-day humans) can pass with facility. <br /><br />It is natural to ponder this question: Had Turing himself been better able to pass the Turing Test of Rorty's essay, might this have helped Turing to find in daily life sufficient <i>hilaritas</i> as to be still with us?John Sidleshttps://www.blogger.com/profile/16286860374431298556noreply@blogger.comtag:blogger.com,1999:blog-3722233.post-35414534752480615982012-01-06T13:57:25.835-06:002012-01-06T13:57:25.835-06:00I wonder how many computer scientists think of Wat...I wonder how many computer scientists think of Watson as interesting AI research.Rahulnoreply@blogger.comtag:blogger.com,1999:blog-3722233.post-68708324953115618632012-01-06T13:28:23.227-06:002012-01-06T13:28:23.227-06:00The problem of integrating concept-driven, rationa...The problem of integrating concept-driven, rational faculties with data-driven, empirical faculties is very old and very deep. I have to agree that I haven't seen much progress on that score in mainstream AI over the last 30 years. Most research traditions tend to drift to one bank of the stream and ignore the other from that point on.Jon Awbreyhttps://www.blogger.com/profile/13378153853941137426noreply@blogger.comtag:blogger.com,1999:blog-3722233.post-66681473064240642772012-01-06T12:46:38.434-06:002012-01-06T12:46:38.434-06:00dip the string in water, freeze it, then push
let...dip the string in water, freeze it, then push<br /><br />let's get humans all good at concepts first :)Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-3722233.post-3200293145307364552012-01-06T12:07:08.895-06:002012-01-06T12:07:08.895-06:00And here I thought the right question was "wh...And here I thought the right question was "which concept classes possess low-complexity polynomial threshold functions." I'm not sure what it means for a computer to understand a concept, but my best definition would involve probabilities.Justinnoreply@blogger.comtag:blogger.com,1999:blog-3722233.post-34632072738677873802012-01-06T12:06:44.728-06:002012-01-06T12:06:44.728-06:00Tell us more about Work War 2.0Tell us more about Work War 2.0Jon Awbreyhttp://jonawbrey.wordpress.com/noreply@blogger.comtag:blogger.com,1999:blog-3722233.post-74663236900125655732012-01-06T11:00:44.557-06:002012-01-06T11:00:44.557-06:00He wants computers to understand concepts not just...He wants computers to understand concepts not just provide "probabilities". One example is why we can pull an object with a string but we can't push an object with a string.Lance Fortnowhttps://www.blogger.com/profile/06752030912874378610noreply@blogger.comtag:blogger.com,1999:blog-3722233.post-87116866039542349932012-01-06T10:28:25.292-06:002012-01-06T10:28:25.292-06:00What did Minsky think the *right* problem was?What did Minsky think the *right* problem was?Justinnoreply@blogger.com