Symmetry-based learning, Domingos, 2014 https://www.microsoft.com/en-us/research/video/symmetry-based-learning/
Approach 2: Deep symmetry networks generalize convolutional neural networks by tying parameters and pooling over an arbitrary symmetry group, not just the translation group. In preliminary experiments, they outperformed convnets on a digit recognition task. On Friday, August 5, 2016 at 4:56:45 PM UTC-3, Kevin Liu wrote: > > Minsky died of a cerebral hemorrhage at the age of 88.[40] > <https://en.wikipedia.org/wiki/Marvin_Minsky#cite_note-40> Ray Kurzweil > <https://en.wikipedia.org/wiki/Ray_Kurzweil> says he was contacted by the > cryonics organization Alcor Life Extension Foundation > <https://en.wikipedia.org/wiki/Alcor_Life_Extension_Foundation> seeking > Minsky's body.[41] > <https://en.wikipedia.org/wiki/Marvin_Minsky#cite_note-Kurzweil-41> Kurzweil > believes that Minsky was cryonically preserved by Alcor and will be revived > by 2045.[41] > <https://en.wikipedia.org/wiki/Marvin_Minsky#cite_note-Kurzweil-41> Minsky > was a member of Alcor's Scientific Advisory Board > <https://en.wikipedia.org/wiki/Advisory_Board>.[42] > <https://en.wikipedia.org/wiki/Marvin_Minsky#cite_note-AlcorBoard-42> In > keeping with their policy of protecting privacy, Alcor will neither confirm > nor deny that Alcor has cryonically preserved Minsky.[43] > <https://en.wikipedia.org/wiki/Marvin_Minsky#cite_note-43> > > We better do a good job. > > On Friday, August 5, 2016 at 4:45:42 PM UTC-3, Kevin Liu wrote: >> >> *So, I think in the next 20 years (2003), if we can get rid of all of the >> traditional approaches to artificial intelligence, like neural nets and >> genetic algorithms and rule-based systems, and just turn our sights a >> little bit higher to say, can we make a system that can use all those >> things for the right kind of problem? Some problems are good for neural >> nets; we know that others, neural nets are hopeless on them. Genetic >> algorithms are great for certain things; I suspect I know what they're bad >> at, and I won't tell you. (Laughter)* - Minsky, founder of CSAIL MIT >> >> *Those programmers tried to find the single best way to represent >> knowledge - Only Logic protects us from paradox.* - Minsky (see >> attachment from his lecture) >> >> On Friday, August 5, 2016 at 8:12:03 AM UTC-3, Kevin Liu wrote: >>> >>> Markov Logic Network is being used for the continuous development of >>> drugs to cure cancer at MIT's CanceRX <http://cancerx.mit.edu/>, on >>> DARPA's largest AI project to date, Personalized Assistant that Learns >>> (PAL) <https://pal.sri.com/>, progenitor of Siri. One of Alchemy's >>> largest applications to date was to learn a semantic network (knowledge >>> graph as Google calls it) from the web. >>> >>> Some on Probabilistic Inductive Logic Programming / Probabilistic Logic >>> Programming / Statistical Relational Learning from CSAIL >>> <http://people.csail.mit.edu/kersting/ecmlpkdd05_pilp/pilp_ida2005_tut.pdf> >>> (my >>> understanding is Alchemy does PILP from entailment, proofs, and >>> interpretation) >>> >>> The MIT Probabilistic Computing Project (where there is Picture, an >>> extension of Julia, for computer vision; Watch the video from Vikash) >>> <http://probcomp.csail.mit.edu/index.html> >>> >>> Probabilistic programming could do for Bayesian ML what Theano has done >>> for neural networks. <http://www.inference.vc/deep-learning-is-easy/> - >>> Ferenc Huszár >>> >>> Picture doesn't appear to be open-source, even though its Paper is >>> available. >>> >>> I'm in the process of comparing the Picture Paper and Alchemy code and >>> would like to have an open-source PILP from Julia that combines the best of >>> both. >>> >>> On Wednesday, August 3, 2016 at 5:01:02 PM UTC-3, Christof Stocker wrote: >>>> >>>> This sounds like it could be a great contribution. I shall keep a >>>> curious eye on your progress >>>> >>>> Am Mittwoch, 3. August 2016 21:53:54 UTC+2 schrieb Kevin Liu: >>>>> >>>>> Thanks for the advice Cristof. I am only interested in people wanting >>>>> to code it in Julia, from R by Domingos. The algo has been successfully >>>>> applied in many areas, even though there are many other areas remaining. >>>>> >>>>> On Wed, Aug 3, 2016 at 4:45 PM, Christof Stocker < >>>>> stocker....@gmail.com> wrote: >>>>> >>>>>> Hello Kevin, >>>>>> >>>>>> Enthusiasm is a good thing and you should hold on to that. But to >>>>>> save yourself some headache or disappointment down the road I advice a >>>>>> level head. Nothing is really as bluntly obviously solved as it may >>>>>> seems >>>>>> at first glance after listening to brilliant people explain things. A >>>>>> physics professor of mine once told me that one of the (he thinks) most >>>>>> malicious factors to his past students progress where overstated >>>>>> results/conclusions by other researches (such as premature announcements >>>>>> from CERN). I am no mathematician, but as far as I can judge is the no >>>>>> free >>>>>> lunch theorem of pure mathematical nature and not something induced >>>>>> empirically. These kind of results are not that easily to get rid of. If >>>>>> someone (especially an expert) states such a theorem will prove wrong I >>>>>> would be inclined to believe that he is not talking about literally, but >>>>>> instead is just trying to make a point about a more or less practical >>>>>> implication. >>>>>> >>>>>> >>>>>> Am Mittwoch, 3. August 2016 21:27:05 UTC+2 schrieb Kevin Liu: >>>>>>> >>>>>>> The Markov logic network represents a probability distribution over >>>>>>> the states of a complex system (i.e. a cell), comprised of entities, >>>>>>> where >>>>>>> logic formulas encode the dependencies between them. >>>>>>> >>>>>>> On Wednesday, August 3, 2016 at 4:19:09 PM UTC-3, Kevin Liu wrote: >>>>>>>> >>>>>>>> Alchemy is like an inductive Turing machine, to be programmed to >>>>>>>> learn broadly or restrictedly. >>>>>>>> >>>>>>>> The logic formulas from rules through which it represents can be >>>>>>>> inconsistent, incomplete, or even incorrect-- the learning and >>>>>>>> probabilistic reasoning will correct them. The key point is that >>>>>>>> Alchemy >>>>>>>> doesn't have to learn from scratch, proving Wolpert and Macready's no >>>>>>>> free >>>>>>>> lunch theorem wrong by performing well on a variety of classes of >>>>>>>> problems, >>>>>>>> not just some. >>>>>>>> >>>>>>>> On Wednesday, August 3, 2016 at 4:01:15 PM UTC-3, Kevin Liu wrote: >>>>>>>>> >>>>>>>>> Hello Community, >>>>>>>>> >>>>>>>>> I'm in the last pages of Pedro Domingos' book, the Master Algo, >>>>>>>>> one of two recommended by Bill Gates to learn about AI. >>>>>>>>> >>>>>>>>> From the book, I understand all learners have to represent, >>>>>>>>> evaluate, and optimize. There are many types of learners that do >>>>>>>>> this. What >>>>>>>>> Domingos does is generalize these three parts, (1) using Markov Logic >>>>>>>>> Network to represent, (2) posterior probability to evaluate, and (3) >>>>>>>>> genetic search with gradient descent to optimize. The posterior can >>>>>>>>> be >>>>>>>>> replaced for another accuracy measure when it is easier, as genetic >>>>>>>>> search >>>>>>>>> replaced by hill climbing. Where there are 15 popular options for >>>>>>>>> representing, evaluating, and optimizing, Domingos generalized them >>>>>>>>> into >>>>>>>>> three options. The idea is to have one unified learner for any >>>>>>>>> application. >>>>>>>>> >>>>>>>>> There is code already done in R https://alchemy.cs.washington.edu/. >>>>>>>>> My question: anybody in the community vested in coding it into Julia? >>>>>>>>> >>>>>>>>> Thanks. Kevin >>>>>>>>> >>>>>>>>> On Friday, June 3, 2016 at 3:44:09 PM UTC-3, Kevin Liu wrote: >>>>>>>>>> >>>>>>>>>> https://github.com/tbreloff/OnlineAI.jl/issues/5 >>>>>>>>>> >>>>>>>>>> On Friday, June 3, 2016 at 11:17:28 AM UTC-3, Kevin Liu wrote: >>>>>>>>>>> >>>>>>>>>>> I plan to write Julia for the rest of me life... given it >>>>>>>>>>> remains suitable. I am still reading all of Colah's material on >>>>>>>>>>> nets. I ran >>>>>>>>>>> Mocha.jl a couple weeks ago and was very happy to see it work. >>>>>>>>>>> Thanks for >>>>>>>>>>> jumping in and telling me about OnlineAI.jl, I will look into it >>>>>>>>>>> once I am >>>>>>>>>>> ready. From a quick look, perhaps I could help and learn by >>>>>>>>>>> building a very >>>>>>>>>>> clear documentation of it. Would really like to see Julia a leap >>>>>>>>>>> ahead of >>>>>>>>>>> other languages, and plan to contribute heavily to it, but at the >>>>>>>>>>> moment am >>>>>>>>>>> still getting introduced to CS, programming, and nets at the basic >>>>>>>>>>> level. >>>>>>>>>>> >>>>>>>>>>> On Friday, June 3, 2016 at 10:48:15 AM UTC-3, Tom Breloff wrote: >>>>>>>>>>>> >>>>>>>>>>>> Kevin: computers that program themselves is a concept which is >>>>>>>>>>>> much closer to reality than most would believe, but julia-users >>>>>>>>>>>> isn't >>>>>>>>>>>> really the best place for this speculation. If you're actually >>>>>>>>>>>> interested >>>>>>>>>>>> in writing code, I'm happy to discuss in OnlineAI.jl. I was >>>>>>>>>>>> thinking about >>>>>>>>>>>> how we might tackle code generation using a neural framework I'm >>>>>>>>>>>> working >>>>>>>>>>>> on. >>>>>>>>>>>> >>>>>>>>>>>> On Friday, June 3, 2016, Kevin Liu <kvt...@gmail.com> wrote: >>>>>>>>>>>> >>>>>>>>>>>>> If Andrew Ng who cited Gates, and Gates who cited Domingos >>>>>>>>>>>>> (who did not lecture at Google with a TensorFlow question in the >>>>>>>>>>>>> end), were >>>>>>>>>>>>> unsuccessful penny traders, Julia was a language for web design, >>>>>>>>>>>>> and the >>>>>>>>>>>>> tribes in the video didn't actually solve problems, perhaps this >>>>>>>>>>>>> would be a >>>>>>>>>>>>> wildly off-topic, speculative discussion. But these statements >>>>>>>>>>>>> couldn't be >>>>>>>>>>>>> farther from the truth. In fact, if I had known about this video >>>>>>>>>>>>> some >>>>>>>>>>>>> months ago I would've understood better on how to solve a problem >>>>>>>>>>>>> I was >>>>>>>>>>>>> working on. >>>>>>>>>>>>> >>>>>>>>>>>>> For the founders of Julia: I understand your tribe is mainly >>>>>>>>>>>>> CS. This master algorithm, as you are aware, would require >>>>>>>>>>>>> collaboration >>>>>>>>>>>>> with other tribes. Just citing the obvious. >>>>>>>>>>>>> >>>>>>>>>>>>> On Friday, June 3, 2016 at 10:21:25 AM UTC-3, Kevin Liu wrote: >>>>>>>>>>>>>> >>>>>>>>>>>>>> There could be parts missing as Domingos mentions, but >>>>>>>>>>>>>> induction, backpropagation, genetic programming, probabilistic >>>>>>>>>>>>>> inference, >>>>>>>>>>>>>> and SVMs working together-- what's speculative about the >>>>>>>>>>>>>> improved versions >>>>>>>>>>>>>> of these? >>>>>>>>>>>>>> >>>>>>>>>>>>>> Julia was made for AI. Isn't it time for a consolidated view >>>>>>>>>>>>>> on how to reach it? >>>>>>>>>>>>>> >>>>>>>>>>>>>> On Thursday, June 2, 2016 at 11:20:35 PM UTC-3, Isaiah wrote: >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> This is not a forum for wildly off-topic, speculative >>>>>>>>>>>>>>> discussion. >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> Take this to Reddit, Hacker News, etc. >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> On Thu, Jun 2, 2016 at 10:01 PM, Kevin Liu <kvt...@gmail.com >>>>>>>>>>>>>>> > wrote: >>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> I am wondering how Julia fits in with the unified tribes >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> mashable.com/2016/06/01/bill-gates-ai-code-conference/#8VmBFjIiYOqJ >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>>> https://www.youtube.com/watch?v=B8J4uefCQMc >>>>>>>>>>>>>>>> >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> >>>>>
