tag:blogger.com,1999:blog-15418143.post6794382104242630346..comments2019-03-25T05:21:10.292-05:00Comments on Tombone's Computer Vision Blog: Deep Learning vs Probabilistic Graphical Models vs LogicTomasz Malisiewiczhttps://plus.google.com/107912691630546731185noreply@blogger.comBlogger9125tag:blogger.com,1999:blog-15418143.post-59316926592638091472016-05-19T09:56:01.011-05:002016-05-19T09:56:01.011-05:00In short, logics are all about proof, while probab...In short, logics are all about proof, while probabilistic/statistical data processing tools/techniques (such as deep neural nets) is all about "exploratory maths" (data clustering, ...). But I really do not agree with the perspectives about logic: it is the calculus of computer science and as such, its foundations...<br />Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-15418143.post-11776817727326009792016-03-19T20:44:21.799-05:002016-03-19T20:44:21.799-05:00Well, What about writing a competing paper to expl...Well, What about writing a competing paper to explain the part that was forgotten instead of "Wft" people. Any point of view is welcome.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-15418143.post-6372007199384422662015-11-02T10:48:56.522-05:002015-11-02T10:48:56.522-05:00> I remember studying skolemization in my AI cl...> I remember studying skolemization in my AI class as an undergraduate. And in my 10 years of robotics research I haven't seen it used once.<br /><br />This is, for the most part, true. Logics are pretty still well used in higher forms of AI than just perception and control. Doug Lenat's work is still primarily logic based. <br /><br />Logic seems to be the win when you're trying to do higher order cognition. It's representation is far more efficient in space and time complexity than these deep methods. <br /><br />I think what you see is that robotics research has not been able to make the case that higher forms of cognition are needed for everyday robotics tasks. ndepalmahttps://www.blogger.com/profile/14975299562460287103noreply@blogger.comtag:blogger.com,1999:blog-15418143.post-39937571839831639542015-10-07T04:05:15.060-05:002015-10-07T04:05:15.060-05:00Considering you majored in Computer Vision, your a...Considering you majored in Computer Vision, your argument is comprehensible.<br />However, it is better to expand your definition about AI, not classification like mnist. <br /><br />I recommend reading 'The Master Algorithm' written by Pedro Domingos. <br />It will give you bird eye view. <br /><br />Pedro Domingos is a leading machine learning researcher.<br />He recently developed "Tractable Deep Learning" ( http://www.cs.washington.edu/node/8805 )<br />His approach is combining first-order logic and probabilistic graphical models in a single representation. <br />Have fun! Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-15418143.post-20794975374051268272015-05-19T02:10:55.539-05:002015-05-19T02:10:55.539-05:00Confusing article: AI is not classification only: ...Confusing article: AI is not classification only: it is just one intermediary part of the AI chain. <br /><br />We miss what is true core AI: knowledge formation, creation of thought processes, rooting of though processes in observed data, interconnection of thought processes over time to form more elaborate ones, along with their harnessing against observed data over time. <br /><br />If you want to start from something practical: create a board scene with items interacting between each other using arbitrary rules and try to write a program with minimal prior knowledge that can understand how the items interact between each other. Have fun! Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-15418143.post-90553560623373683642015-05-14T15:04:42.197-05:002015-05-14T15:04:42.197-05:00I need to agree with Anonymous, here.
Your post r...I need to agree with Anonymous, here.<br /><br />Your post repeats a popular narrative -- one that's simplified to the point of being completely wrong. Logic programming isn't a casualty of the AI winter. Furthermore, first order logic is not representative of the variety of logical systems in common use -- it's merely the easiest one to teach to freshmen.<br /><br />Logic programming has clear benefits over statistical methods for particular sets of applications. After all, where statistical methods are completely opaque, the equivalent logic is typically human-readable. And, of course, logic programming continues to dominate in the tasks where it has always been dominant -- expert systems, because it makes no sense to try to get a model to learn patterns when a human being has already learned the patterns and can trivially codify them; theorem proving and constraint solving, because the starting point is formal and the inferences are formal.<br /><br />Probablistic methods are not probablistic logics. This seems to be a blind spot in your background -- a probablistic logic is a system that combines logic-style inference rules with fractional truth values, thus combining the expressive clarity of logic with the expressive completeness of probability. (A good example is PLN, which Goertzel has written on.)<br /><br />Certainly, statistical methods are easier to use when producing 'theory-free' models empirically. But, we rarely want to be truly 'theory-free' -- doing so means putting us at the mercy of obvious absurdities produced by biased samples. Logic programming has its place performing automatic sanity checks, and flagging situations to the developer when the statistical model strays outside the bounds of what some logical and human-readable model considers reasonable.<br /><br />Logic also has its place in didiacticism. After all, logic is readily comprehensible to students.<br /><br />Certainly, nobody could reasonably make the case that logic programming should fully displace statistical models! But, it's equally unreasonable to suggest that statistical models should be used when logic programming can do it better. And, to suggest that formal logic should be consigned to the past is to ignore the actual situation in industry and in academia in favor of a misunderstanding of history: statistical models did not rise in popularity because of a growing failure of logical models, but because logical models operate on a human scale while statistical models operate on an inhuman scale, and so statistical models gain popularity as the inhuman scale at which they must operate becomes more practical. The set of situations wherein statistical models are displacing formal models is the set of situations wherein formal models would have never been used, had resources not been a factor.John Ohnohttps://www.blogger.com/profile/11352441770252592928noreply@blogger.comtag:blogger.com,1999:blog-15418143.post-79875866637759808872015-04-20T05:07:19.632-05:002015-04-20T05:07:19.632-05:00Nice website!Nice website!patriciahttp://www.arcticsecrets.nl/noreply@blogger.comtag:blogger.com,1999:blog-15418143.post-79852145533435105112015-04-12T10:25:59.440-05:002015-04-12T10:25:59.440-05:00When I mention logic in the blog post I am really ...When I mention logic in the blog post I am really referring to First order logic.<br /><br />I remember studying skolemization in my AI class as an undergraduate. And in my 10 years of robotics research I haven't seen it used once.<br /><br />Probabilistic methods are broad and this blog post is actually defending those methods over black-box deep learning methods.<br /><br />We don't use the term logic the way you'd be happy with. We don't call probabilistic methods "probabilistic logics."<br /><br />Feel free to post links to some of your published papers and/or articles. And thanks for the insightful comment.<br />Tomasz Malisiewiczhttps://www.blogger.com/profile/17507234774392358321noreply@blogger.comtag:blogger.com,1999:blog-15418143.post-45232737893232853922015-04-12T07:17:08.486-05:002015-04-12T07:17:08.486-05:00Apparently, you know nothing about logic-based AI....Apparently, you know nothing about logic-based AI. Logic has long advanced beyond what you describe ("modus ponens", wtf??) - there are probabilistic logics, statistical-relational learning, neuro-symbolic representation, stochastic logic, first-order graphical models and many other logic-based approaches which allow for statistical learning, dealing with partial observability, uncertainty and noise, etc. They have been successfully applied in various real-world areas such as biomedical research and large-scale social network analysis. I suggest you make yourself familiar with the state of the art in research before writing rubbish. Anonymousnoreply@blogger.com