On Mon, Sep 30, 2013 at 9:16 PM, Aseem Hegshetye <axh118...@utdallas.edu>wrote:
> > Hi, > For spatial representation of an input to the cla classifier, an input is > converted in 121 bits ,21 out of which are on. > Every column is connected to 50% of the bits. And with the range of min to > max, ON bits are distributed from left to right > in those 121 input bits. > That is all i know about conversion of normal input into a spatial input > for cla. I would like to know more details please. > What you describe is how the scalar encoder converts numbers into a binary input for the spatial pooler. If you want more details on what happens in the spatial pooler and the temporal pooler, please read the white paper: http://numenta.org/cla-white-paper.html > > Also Scott purdy said he would reply to my video > http://www.youtube.com/watch?v=7PGFivWS3Kk on CLA classifier with details. > But i havent heard from him. I thought on others question about the > connection matrix predicting the next output. I was messed up designing > something new but CLA spatial pooler claims to produce same sparse > distribution for one type of input. So for same input, same columns in the > CLA would get activated. So with those columns connecting 50% of the input > bits and knowing that 21 total input bits are going to be ON , most > probably same columnar activation should produce the same predicted output ! > Scott purdy said it has been discussed in grok before, and also tried out > may be. > I am very curious to know about it. > Would it be good for me to try programming it to see if it works. > I recently have learned python to get compatible with Numenta ! > With respect to the video, and as I mentioned before, one of the ideas is similar to reconstruction. We no longer use reconstruction because the CLA classifier gives better results. There is no biological equivalent to these components so it is more a matter of what gives the best results for your problem. The other idea was about predicting multiple steps in the future. In some ways it seems similar to temporal pooling if I understand it correctly. Right now we convert the predicted TP state into predicted values using the CLA classifier and that works pretty well. Hierarchy and temporal pooling can help for certain types of problems but we don't currently use them. It would be great if you experimented with your idea - I would be curious to see how it works. > > Thanks > Aseem Hegshetye > > > _______________________________________________ > nupic mailing list > nupic@lists.numenta.org > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org >
_______________________________________________ nupic mailing list nupic@lists.numenta.org http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
