The temporal memory internally contains a bunch of different states such as active cells and predictive cells. The TM output is typically the indices of the active cells, and this is fed to the next level (such as CLA Classifer).
The anomaly score is computed using the previously predicted columns of the TM and the current active columns of the spatial pooler. --Subutai On Sun, Oct 18, 2015 at 9:23 PM, cogmission (David Ray) < [email protected]> wrote: > Hi Subutai, > > Please clarify. The input is a sparse vector of the indexes of predicted > columns correct? Namely the TM output? > > On Sun, Oct 18, 2015 at 11:13 PM, Subutai Ahmad <[email protected]> > wrote: > >> The CLA classifier normally gets the active cells of the TM passed in >> (not active columns or predicted cells/columns). >> >> --Subutai >> >> >> On Oct 18, 2015, at 21:01, cogmission (David Ray) < >> [email protected]> wrote: >> >> Hi Sebastian, >> >> I'm not sure what you were doing, but as far as I've always known, the >> CLAClassifier takes the TM output of predicted columns? There term "Active >> columns" usually refers to the SP input. >> >> It's confusing because the CLAClassifier method documentation says >> "Active Columns" --> >> >> patternNZ: list of the active indices from the output below >> >> >> This should probably be changed to "predicted columns". The clue that I >> take is that the column indices passed in are "sparse" (meaning that the >> input vector consists of the indices of columns which have on bits). (i.e >> [1,3,5] instead of [0,1,0,1,0,1]) >> >> Cheers, >> David >> >> On Sun, Oct 18, 2015 at 10:35 PM, Sebastián Narváez <[email protected]> >> wrote: >> >>> Regarding my previous question, I've encountered that the error was >>> rather silly. I was feeding the CLA Classifier with the predicted columns >>> instead of the active ones. Changing that turned out in accuarate >>> multi-step predictions. >>> >>> On Sun, Oct 18, 2015 at 3:31 PM, Sebastián Narváez <[email protected]> >>> wrote: >>> >>>> Hey guys. I've trained nupic with some data, but when I try to predict >>>> the outcome given an input with the CLA Classifier, it always predicts the >>>> output belonging to the first sequence. I've tried changing the order of >>>> which I pass the data to Nupic, but it will just predict the output of the >>>> new first sequence. Am I doing something wrong? >>>> I've attached the data I'm using for these tests. For each column, each >>>> word is passed to a TM. Then the Active cells of both are passed into a >>>> general TM (first all the words from the first column, and then all the >>>> words from the second column). I reset() all the TMs for each row. >>>> >>>> Thanks in advance. >>>> >>> >>> >> >> >> -- >> *With kind regards,* >> >> David Ray >> Java Solutions Architect >> >> *Cortical.io <http://cortical.io/>* >> Sponsor of: HTM.java <https://github.com/numenta/htm.java> >> >> [email protected] >> http://cortical.io >> >> > > > -- > *With kind regards,* > > David Ray > Java Solutions Architect > > *Cortical.io <http://cortical.io/>* > Sponsor of: HTM.java <https://github.com/numenta/htm.java> > > [email protected] > http://cortical.io >
