Hi Nick, just two thoughts: The SP does not perform well when the number of columns is too low (hence the default 2048), maybe you have too few categories and this is the reason? FOr classification, you could use UtilityEncoder that I;ve merged recently. Append label to your training data (with the evaluationFn) and then pass the testing data and SP should "reconstruct" the most probable label to that sample.
cheers, On Tue, May 13, 2014 at 7:04 AM, Chetan Surpur <[email protected]> wrote: > Nick, > > I don't have an answer to your particular problem, but I want to point out > a resource that may help. > > It's new and experimental, but Cerebro 2 will let you see the state of > your model at each point in time, including the connections in the SP. > https://github.com/chetan51/nupic.cerebro2 > > Perhaps you can use it to see more clearly what the SP is doing. Let me > know how it goes! > > - Chetan > > On May 12, 2014 at 6:42:19 AM, Nicholas Mitri ([email protected]) wrote: > > Hey all, > > So I’ve been running clustering tests on the SP and comparing it against > other algorithms. > > I’m passing the data points through the SP for several passes while > restricting the number of columns to that of the desired cluster number. > The results, unfortunately, seem to be very inconsistent. Using the RAND > index to evaluate the outcome relative to the true labels of the data > points, I’m getting results varying between 56% and 95% on one of Wine > database for example. > > Where is this inconsistency coming from? I‘ve set the potential pool to > span the entire input with potentialPCT set to 1. I thought that would > reduce the effect of the random permanence initialization step. It > evidently hasn’t. > > Any tips are appreciated! > > best, > Nick > _______________________________________________ > 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 > > -- Marek Otahal :o)
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
