Hi Stefan, Subutai is correct, but it's perhaps not clear what he's saying.
The first time you run the entire sequence in your diagram EVEN|EVENEVEN (TP reset marked by |), the cells in E which are chosen are different depending on which E you're talking about. Let's say for simplicity they're chosen in sequence from the top in each column. (a) E is active and bursts. E1 -> V1 (E1 happens to predict V1) (b) V is active and V1 is chosen. E1 -> V1 is learned. V1 -> E2 (c) E is active and E2 is chosen. V1 -> E2 is learned. E2 -> N1 (d) N is active and N1 is chosen. E2 -> N1 is learned. N1 -> E3 reset. (e)-(g) is a repetition of (a)-(d). (h) E is active and E3 is chosen. N1-> E3 is learned. E3 -> V2 (i) V2 -> E4, (j) E4 -> N2 So, the CLA has learned EVENEVEN, but the second EVEN is different to the first. Once the CLA has learned the sequence, this happens: (a2) E is active and bursts. V1, N1, V2, N2 predicted. (b2) V1 is active and predicts E2, and so on. The sequence of activation the second run through is thus: E* V1 E2 N1 | E* V1 E2 N1 E3 V2 E4 N2 When Subutai says that this may require several repetitions, this is because all these "happens to" connections would actually need to be learned as a result of seeing the transitions many times, and possibly (or more correctly, probably) involving connections spanning the many active columns for each letter. For simplicity, you've represented each character as one column, but in a real CLA there will be forty active columns, so the above (a)-(j) will be happening using possible combinations of 40x40 columns for each step! So 40 E columns burst in (a), and potentially 40 V1's will become predictive. Initially just some of the V1's will do so, but all of them will raise their permanences to whichever E1's they're connected to. Over several repetitions, these transitions will thus be learned. If, in addition, you have an SP which adds in predictive potential (as the theory recommends..), then you will have the V columns which are predictive chosen ahead of those which aren't. This will allow for better learning. Regards Fergal Byrne On Fri, Dec 13, 2013 at 10:07 AM, Stefan Lattner <[email protected]>wrote: > Hi Subutai, > > thank you very much for your answer. Apparently I do not understand > correctly, when a cell is considered to be predicted and should be re-used > when the sequence occurs again. > > If I understand your diagram correctly, picture (h) is the key one. In > this case you want a different cell in V to become active. The first time > through (h), when you see E that column bursts and the old V cell will > become predicted. The next time through though E will not burst and the > cell in E you show as L will get predicted. Then when V happens, it will > burst because it has not been predicted and a new cell in V will get chosen > to be the learning cell. And so on. > > > Something is not clear for me, this is how I understand it: > If the next time the sequence occurs, "the cell in E I show as L is > predicted", it will get active at the next time step, because E actually > occurs. This activation will cause the cell in V to be predicted and > therefore the column will not burst. You meant it will burst, because it > would not be predicted. I thought, that a cell becomes predicted as soon as > enough cells are active which have a forward connection to that cell. With > desired local act of 1, it would be enough if the cell in E is active and > predicts the cell in V. Please tell me where I am wrong. > > Thank you again and all the best, > Stefan > > > A lot depends on how you set the parameters, but I believe this is the > correct intuition. > > In NuPIC, you can look at the (admittedly complex) test script > $NUPIC/examples/tp/tp_test.py In there we construct pretty high order > sequences such as: > > A B 0 1 2 3 4 E > G H 0 1 2 3 4 I > > In this example, in order to make the correct predictions for the last > element, the TP has to learn a very high order sequence. To learn this the > TP will likely have to see several repetitions. > > —Subutai > > > > On Wed, Dec 11, 2013 at 10:58 AM, Stefan Lattner > <[email protected]>wrote: > >> Hey guys! >> >> I am new to this list, just found out that it even exists. >> >> Since the white paper on the CLA was published in 2011, I am working on a >> JAVA implementation of that model. >> Because there was no such in NuPIC in 2011, I wrote everything from the >> scratch. Many questions came up during this process but there was no >> detailed description on how it works the white paper was the only resource >> I had. >> >> That is kind of sad, I would have had many questions. I am currently >> finishing my masters thesis where I am writing about everything I found out >> about the HTM and CLA during my experiments, because I didn't expect >> Numenta to come up with more details (which is actually the case, except >> the code Numenta is providing). >> >> A very important question for me is still that, according to Numenta, the >> CLA should have a variable Markov Order. However, the way learning is >> described in the CLA white paper leads to an order of two only. >> >> I made an illustration for my thesis how learning takes place, this is >> how I understand it. >> (At first, the sequence E-V-E-N is learnt (Desired local activity = 1). >> Then, the HTM is reset and E-V-E-N is provided to the HTM again. Then, the >> HTM is not reset while E-V-E-N is provided again). >> Can somebody tell me how the order of learnt chains can get increased or >> what I am understanding wrong? >> >> Thank you and all the best, >> Stefan >> >> <temporal_pooler.png> >> >> _______________________________________________ >> nupic mailing list >> [email protected] >> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org >> >> > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > > > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > > -- Fergal Byrne, Brenter IT <http://www.examsupport.ie>http://inbits.com - Better Living through Thoughtful Technology e:[email protected] t:+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
