While it may turn out (due to everyone starting off their models with the same random number generator seed) that the same sequence of data will lead to the exact same SDR being generated in the SP, this is a pathological side-effect of the design, and is only useful for testing the code.
The purpose of the SP is to learn how to best map the input data to SDR's. Which exact set of columns appear in a single SDR is irrelevant for non-topological data, because the input bits are *by design* randomly assigned to columns by the algorithm. For topological data, this random assignment is repeated, with the restriction of the choice of input bits to a topological sub-region. In addition, the input data should always be semantically redundant, with closely related inputs having significant overlap. This should in turn give rise to SDR's which have overlapping active columns to the same extent. Exact matches are not as important as close ones, because the SP is learning the texture of the semantic terrain rather than relying on discrete sampling. The input to all real world systems (including the neocortex) is noisy and incomplete, and the task here is to find the structure embedded in the data despite this noise. So the SP will by necessity have a robust evolution which does not depend on any single input datum, but gradually accommodates the data as it comes in. Regards, Fergal Byrne On Sun, Sep 15, 2013 at 5:25 AM, Hideaki Suzuki <[email protected]> wrote: > Hi, > > I'm not reading this thread in depth, so if I say something inappropriate > in this context, please forgive me... > > I believe NuPIC's SP is completely deterministic, IF you think of the > temporal changes in the input pattern. When you feed the same input > pattern "sequence" it will give you the same SDR always. No random > generator necessary within SP. > > This is because SP tries to best represent the input patterns with the > usable amount of memory (the number of columns). If you provide only one > input pattern, it tries to create SDR using all columns to represent that > single pattern (you may end up with several SDRs to mean the same single > pattern). If you provide 100 patterns, the columns will get segregated by > themselves so that the region can convert all 100 patterns to 100 SDRs with > its best accuracy and usage of all columns. If some columns become > defunct, then the HTM region changes its SDR (e.g. using some least used > columns to replace the dead ones) to again best represent the input > patterns. > > Everything is autonomous. > It is a beautiful algorithm to me. > > Considering pattern change in time during creating sparse representations > might be one of the unique features of HTM CLA. (I'm not sure because I'm > not a neuroscientist). > > In short, you can get the same representations by either of two ways. > 1. always feed the same input pattern sequence. > or > 2. always reset SP after feeding one (i.e. disabling learning) > > Best Regards, > Hideaki Suzuki. > > > _______________________________________________ > 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
