On Fri, Jul 11, 2014 at 11:04 PM, Alexander Hirner <[email protected]> wrote:
> Great test case. As far as I understood, you build separate models for > each frequency bin and aggregate a total anomaly score by an arbitrary > number of how many single anomaly score must be above a certain treshold. Correct. And the # of anomalyLikelihood values that cause the trigger is configurable, as well as the anomalyLikelihood threshold itself. You can do a lot of tuning with these two values, which I haven't done much of yet. I'm still trying to find the best parameters for these for the music I'm passing in. > What do you think about dumping all the bins into one TP at once? In that > way, the model could also capture covariance between the frequency bins and > derive on a conclusion of an aggregate anomaly score on its own. > I wouldn't know how to do that. The input bin values would need to be converted into SDRs, which is what the SP does. The only time I've ever fed data directly into the TP is when my data was already in SDR format (from Cortical.IO API). --------- Matt Taylor OS Community Flag-Bearer Numenta
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
