Hey Doug, Yes, profiling is definitely part of the process typically but I chose to emphasize actual clustering performance here. CLA isn’t intended for clustering but its pooling mechanisms allow it to do so in a limited capacity. Largely due to the binning being performed by the encoders, I think it’s mostly capable of clustering data with "city-block” clusters. More to the point, the end effect is more vector quantization than actual clustering. In the latter, groups of data points in all sorts of spatial configurations are ideally grouped together as long as there’s evident locality within a cluster and separation from others. This is why the hierarchical clustering alg described in the document performs so well. Expectation maximization clustering algs perform greatly here too. As for CLA, it can be used as a quasi-clusterer.
In terms of processing speed, it is true that its not the most efficient given the recommended size of the regions (2048 columns). For the experiments in the doc, I used an 8-by-8 region instead and it performed much better for the synthetic clustering data. At this size, the speeds of convergence were very comparable. Unfortunately, I didn’t do any profiling to measure those quantitatively. Thanks for your feedback :)! best, Nick On Apr 18, 2014, at 9:30 PM, Doug King <[email protected]> wrote: > Well done Nick, > > I have a question about how much processing is used to get results. It's > arguable that to evaluate the performance of a particular clustering algo one > would need to include processing cycles consumed. I realize that the CLA > method is particular to the CLA and needs to use columns, dendrites, etc. > operate, so perhaps is not as efficient. > > Cheers, > > Doug > > > On Thu, Apr 17, 2014 at 7:41 AM, Nicholas Mitri <[email protected]> wrote: > Hello all. > > I just completed a rough draft (and by rough I mean rough!) of a document on > evaluating the spatial pooler as a clustering algorithm. > I’m attaching the document here for your thoughts or in case anyone is > interested. > Please excuse any inaccuracies or typos. I’ll refine it on the second pass > before considering adding it to my thesis. > > 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
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
