Hi Subutai, Not sure if you made it to the bottom of my last (big long) post. Eventually I found that CLA does work properly in this case. For some reason, the anomaly score was returned as "nan" and that threw off the plotting. I sent my files to Matt for forensics as he requested. Explicitly searching and replacing nan's with 1.0's made my plots look much better.
Cheers, Ritchie > Message: 1 > Date: Mon, 28 Apr 2014 09:10:33 -0700 > From: Subutai Ahmad <[email protected]> > To: "NuPIC general mailing list." <[email protected]> > Subject: Re: [nupic-discuss] Noisy step example > Message-ID: > < > ca+zatih1krt3khyyhk5pgkp_rnpnxtg6s+g9qif7q98-0h3...@mail.gmail.com> > Content-Type: text/plain; charset="iso-8859-1" > My guess would be the same as Marek's. Or perhaps some other parameter > differences. I'm not sure why the anomaly score doesn't spike at the jump > in your test. We see this type of transition all the time in Grok and it > seems to work well. In Grok we do have learning on always - I don't know if > that could be a difference. > --Subutai Ritchie Lee Research Engineer Carnegie Mellon University-Silicon Valley NASA Ames Research Center Bldg 19, Rm 1024 Moffett Field, CA 94035 (650) 335-2847
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
