To quote Subutai from another thread [1]: "You might not get an exact match unless the inputs are really really close. In particular, the spatial pooler won't do too well at learning general invariances (and it's not supposed to). For example, if you shift the image by one or two pixels you might get a very different output SDR."
[1] http://markmail.org/message/dytllixwodm5bjcu --------- Matt Taylor OS Community Flag-Bearer Numenta On Sun, Mar 9, 2014 at 12:27 PM, Traun Leyden <[email protected]> wrote: > > I created a simple example to recognize digits using the spatial pooler, > it's available on github here: > > https://github.com/tleyden/nupic-digitrecognizer > > It was able to get exact matches with SDR's, and I didn't need to use the > KNNClassifier. Having said that, the test data was relatively "easy", in > that I just went through the training data and removed a few pixels. > > However when I tried to recognize a digit that's been entirely shifted to > the right by a few pixels (original - shifted) it wasn't able to find a > match. > > What are some suggested approaches to add spatial invariance so it can > handle this? > > I did see this mailing list thread but so far no simple answer has jumped > out at me. > > Also thanks to all the folks who responded to my questions, those were > useful pointers. > > > > _______________________________________________ > 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
