http://www.ibm.com/blogs/think/2016/01/14/the-next-grand-challenge-computers-that-converse-like-people/
Jim Bromer On Thu, Jan 14, 2016 at 10:20 AM, Jim Bromer <[email protected]> wrote: > I watched the Presenti presentation on youtube a few days ago. > > Neural Networks can learn but they cannot use that learning efficiently in > many important ways. Discrete AI can acquire more specific > (discrete) 'objects' as they learn. So back in the 90's people started > using hybrids that combined neural networks with discrete methods. Machine > learning includes advances on hybrid methods. > > Most discrete methods are built around networks of relations between the > data objects which represent 'concepts' or 'ideas' or 'knowledge' or 'know > how' or whatever it is that you want to call the data objects that would be > used to hold knowledge in a (more) discrete AI program. So a > contemporary discrete AI program is also going to be an implementation of a > network. The network may include numerical values but even if it doesn't > it probably will represent categories of association. That definition is > not meant to be complete because I am only trying to get an idea across: > Modern discrete AI methods involve network methods that can potentially be > seen as representatives of 'thought' that are more sophisticated than > neural networks. That makes sense. > > Presenti was talking about an IBM researcher from the 70s who found that > he could use statistical methods to *learn* about speech without a > linguist. That would be a form of machine learning. Therefore it is fairly > safe for me to conclude that Watson used machine learning in what Watson > researcher's called, "Deep NLP". > > My question was why haven't there been clear advances in search engine > technology in the 2 years since Deep Learning and Watson have made very > obvious advances in AI? I did an image search for "cats" on google and it > was very good. I only found one dog (a small dog which had been photo > shopped with multiple legs somewhat like a caterpillar.) I tried some other > searches on images and the results were also very good. The results were > really amazing. So there have been some advances on image searches in the > past 2 years. The search for "castles on the moon" did not distinguish > between castles pictured as being on the moon from castles with the moon in > the scene. So even though I am nit-picking to some extent the point is that > it looks like you have to train a deep learning neural network with a > narrow training sample in order to teach it to recognize something that > would require a little thinking outside the box. That was also a problem > with Watson. Its Deep NLP could be trained with all the questions from past > Jeopardy shows (and Jeopardy-style questions that researchers could create) > but can it be trained to handle juxtapositions of linguistic 'concepts' > that might require some thinking outside of the box? (Incidentally I tried > "cat in a box" and Google did very well. But when I tried "full stadium" it > did include pictures of stadiums that were not empty. I could spot them as > I was paging quickly through the images.) But I guess I there have been > some significant advances in the past 2 years. They just do not include > using language to refine your searches. > > My idea of Concept Integration is that different concepts cannot always be > merged, as in a neural network to take an example, because as more concepts > are integrated the requirements of a part of the conceptual integration may > change. To restate that in another way,the integration of a number of > concepts will typically change if additional concepts are integrated with > them. This is what would happen if you tried to refine your search using > conversation. > > Jim Bromer > > > On Tue, Jan 12, 2016 at 10:21 PM, LAU <> wrote: > >> OK, as you wish ... It's just a word. We do not agree on the >> signification. But, it's OK. If you call it "deep learning", or "conceptual >> learning" ... or "Hakuna Matata's learning", it's not important. Stop >> playing with words. >> >> Try to back to the topic of this thread. If I understand what you want to >> promote : >> 1) You note that deep learning implemented in the industry is not so >> intelligent than espected taking in account the computation power available. >> 2) Watson seems to less "narrow" than other implementations. >> 3) What it miss there is "conceptual integration". >> Correct me if I'm wrong. >> >> >> In my humble opinion, there's no intelligent machines just because people >> don't try to, or most likely don't figure out how to, make it more >> intelligent. >> Implementing"conceptual integration" is certainly a way that some >> researchers tried, but lead to no significant results until now. If I look >> at wikipedia, the theory is dated from 1990s. Twenty years later, still >> nothing. >> >> There's no magic behind deep learning, I mean neural networks, used by >> Google or Facebook. Very roughly, it's just a kind of "universal >> approximator". And it's not the computation power that with make it >> spontaneously more intelligent. >> Deep learning becomes very popular these last years because it's easier >> to make a neural network to accomplish picture or voice recognition task >> (*I've >> made a small one myself from scratch in few days*) than handcrafted >> codes, and for a better result. >> But basically, a neural network is just another kind of programming. >> Instead of coding a multitude of operation to achieve a complex task, a >> neural network can do it itself by learning from examples. >> And the question will be how to teach a neural network what is >> "conceptual integration" ? >> >> In the paris tech conference video (*on youtube, but it's in french ...*), >> Jerome Pesenti said something else interesting. He cite a IBM's 70s >> researcher, Fred Jelinek, who said "*Every time I fire a linguist the >> performance of the speech recognizer goes up*". The Jelinek speech >> recognizer team was composed by part of linguist and engineers. By >> replacing a linguist who treats language as a human do by an engineer who >> does mathematics and statistics on words, the result is better. And it >> seems to be the philosophy at IBM to work differently that a human do, and >> it seems to give better result. Instead of playing jeopardy in human way, >> watson applies statistics on the database (*which was wikipedia*). >> >> What I want to say is that may be the "conceptual integration" is a track >> to explore for building AGI. Or, may be the solution will come from >> elsewhere. >> >> >> LAU >> >> >> >> Le 12/01/2016 10:46, Jim Bromer a écrit : >> >> Deep Learning is Deep Machine Learning and Machine Learning is in no >> way limited to Neural Networks. So there is no way that Deep Learning >> is going to be forever defined to refer Machine Learning that uses >> Neural Networks (in certain ways). From that point of view I can say >> that Watson-Jeopardy probably did use a kind of deep learning. >> >> >> ------------------------------------------- >> AGI >> Archives: https://www.listbox.com/member/archive/303/=now >> RSS Feed: https://www.listbox.com/member/archive/rss/303/27172223-36de8e6c >> Modify Your Subscription: https://www.listbox.com/member/?&; >> Powered by Listbox: http://www.listbox.com >> >> >> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >> <https://www.listbox.com/member/archive/rss/303/24379807-653794b5> | >> Modify >> <https://www.listbox.com/member/?&;> >> Your Subscription <http://www.listbox.com> >> > > ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
