Hi Ivan, I'm not sure what you're driving at here, but let me just try to answer your concerns.
While a huge amount of (largely fruitless IMHO) work has been done in classical AI research on NLP, this work has a fatal flaw. It is looking at the problem as one which can be solved by combining ever more complex programs in ever more complex ways. The only good thing it has achieved is that it has asked a lot of questions about the structure of language, and also allowed computers to analyse huge amounts of text in order to get good statistical understandings of what we actually say. Loads of clever software cannot be the answer. Only one system in the universe has been built which does NLP, and that system uses no programs. It has some inbuilt structure, but it's essentially a set of simple CLA's wired up in a particular way. So, what we're attempting to do here is to assess what happens when you look at one (hypothetical) piece of this system. All regions in our language systems communicate by passing SDR's up the hierarchy. So we've got some of these from CEPT and we're going to do some experiments to see can we observe anything similar to what we see in the brain. The classifiers we use are just implementation artefacts which allow us to measure and observe the activation patterns and predictions that the CLA is generating. These classifiers are like electrode arrays or fMRI pictures, and are not part of the simulation. We aren't going to be writing classifiers to perform the tasks you listed - in fact the opposite is the case. We plan to build hierarchies of CLA's which perform those tasks, with only very simple (but massively parallel) plumbing between them providing the "cleverness". The key to figuring out how the brain achieves all these wonderful feats is to figure out how to connect the regions and combine them together in the hierarchy. The research begins with this simple question: How do we (or just can we) connect the CEPT SDR's, representing words, to a "word-level" region of CLA, in order to produce some meaningful linguistic performance? The answer will be dependent on both a) whether the CEPT data is a good analogue for how words are passed around in the brain, and b) whether we can design a way to connect up the CEPT SDR's to the CLA which works. We know neither of these for sure, but by varying the connection strategies, we should be able to provide an answer to the former question. This should allow us (if we fail) to improve the CEPT data and try again. It is possible that we're barking up the wrong tree with the CEPT SDR's. If so, we should find out as soon as we can, and then see if a pivot will yield fruit. Humans are doing something like this, so let's see if we can generate the appropriate analogue with just a single region, or if we need to add hierarchy to get there. Regards Fergal Byrne On Sun, Sep 15, 2013 at 9:12 AM, Ivan Sytenko <[email protected]> wrote: > Hi Fergal, > > Thank you for such a comprehensive answer. > > I am not questioning the value of the semantic information. It seemed to > me that the use of CEPT information in this form will be confused levels of > knowledge representation. You get the semantic information at the > morphological level, where it will hang. While abstract information should > settle on the upper levels to take full advantage in the > deductive-inductive analysis. > > This problem is similar to the transfer of knowledge between the two > stochastic systems. I think that it is closer to the noise. The semantic > information must be obtained directly in your system by the analysis of > texts, or you need another CEPT data format. > > Of course, the CLA has huge potential. But you have to be prepared to > write "classifiers" for new tasks: invariant pattern recognition, planning, > reasoning.. I do not see any CLA mechanisms to solve these challenges. > > As for the initial question, it is interesting to consider the close > association of CLA and CEPT concepts. > > I hope that constructive discussion will help the project. > > Regards, Ivan > > On:Date: Sat, 14 Sep 2013 12:18:00 +0100 > From: Fergal Byrne <[email protected]> > To: "NuPIC general mailing list." <[email protected]> > Subject: Re: [nupic-dev] Subject: Re: HTM in Natural Language > Processing > Message-ID: > <CAD2Q5YfgC_dSGv19yPpFnvVf6HzoAQZNFnmYAr5DyOHVBH3= > [email protected]> > Content-Type: text/plain; charset="iso-8859-1" > > Hi Ivan, > > You're completely correct about increasing the amount of information coming > in, but that is precisely what we want. > > If you just want to treat words as abstract labelled symbols, then we > already have methods for that. In that case all your doing is learning the > sequence of objects, without any "understanding" about *why* the sequence > is as it is. You could just as easily be learning sequences of notes (and > not learning anything about the rules or "quality" of music), or, as we > used to watch on BBC's Generation Game, a sequence of consumer goods > passing by on a conveyor belt. Each successive object gives the CLA no > information about what the the next object should be. > > When you add all this semantic information using the CEPT data, you're > learning what *kinds of words* fit together in a sequence. Humans learn > language based on categorisation, inference and generalisation, so the > stream of words must contain structural information which allows this to > happen. > > For simple English sentences, for example, the first object is usually a > noun (the subject of the sentence), and the second object is usually a verb > which agrees in number with the subject. If the verb is transitive, this is > usually followed by another noun, the object of the sentence. Variations on > this pattern exist, but each one operates on its own very strict set of > rules, again providing a structure which is contained in the (generalised) > grammar of the language, the details of which are contained in the > experienced streams of words. > > As we learn English (or any language), we learn the semantic category and > grammatical settings for each word along with the meaning and sound of the > word. We also learn the rules for fitting these things together in > grammatically correct and semantically sensible ways. We can generalise and > innovate based on the constraints of these rules, and we can detect and > error-correct when the rules are broken. The language itself is learned in > this holistic, holographic way, extracted only from what we hear and learn > to think and produce (there is likely to be some hardwired "Universal > Grammar" in our brains which is probably implemented by the way that > various regions are wired together). > > Pinker's *The Language Instinct* is a great survey of these ideas. > > So, this exercise is a first set of steps in exploring how we can use the > CLA to interface with this semantic structure of natural language. We're > going to see how well a single layer, single region, small CLA deals with > the structural and semantic information about sentences which the CEPT > encoding provides. > > By the way, this is also a test of the usefulness of the CEPT concept and > implementation, which itself was motivated by Francisco's awareness of the > work Jeff et al have been doing on the CLA. We'll be seeing how well the > CEPT data embodies the kind of semantic information we believe is needed > for use in NLP. CEPT have already begun adapting their software in response > to feedback from the NuPIC/Grok side, and they're clearly very interested > in using our work and the hackathon to improve the CEPT SDR system and its > power as an NLP encoding scheme. > > In terms of capacity, there's not much to worry about, IMHO. The CEPT SDR's > are 16k bitmaps, and a 2k CLA should be able to deal with that. The input > here is really an emulation of a (hypothetical) high-level language > processing region being fed with a lower- (but still high-) level data > stream in the language of the neocortex - an SDR. The 16k of data is really > a set of sub-SDR's, each of which is encoding a different semantic aspect > of the underlying word. This is similar (structurally) to the hierarchical > organisation which we know exists in the brain for handling language, and > indeed for most corticocortical hierarchy. > > We'll find out how well this works when we start the exploration, and we > will easily be able to up the size of the CLA if we want to try that. > Personally, I think that a 2k CLA has an enormous capacity - much higher > than you would think - because of the way that it efficiently deploys > connections and predictor cells based solely on the actual content and > structure in the data. > > Regards, > > Fergal Byrne > > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > > -- Fergal Byrne ExamSupport/StudyHub [email protected] http://www.examsupport.ie Dublin in Bits [email protected] http://www.inbits.com +353 83 4214179 Formerly of Adnet [email protected] http://www.adnet.ie
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
