Hi, Subutai, Thanks for the detailed reply. Really appreciate it! :)
Regards, Omair S > On Jul 28, 2014, at 9:42 PM, Subutai Ahmad <[email protected]> wrote: > > Hi Omair, > > Thanks! We have only done rough comparisons so far with standard models. One > problem has been the lack of good labeled datasets for streaming anomaly > detection (which is the specific application we've tuned for). We have been > accumulating some real world data and hope to put together a dataset for more > rigorous testing. I hope to release that soon. If you know of some good > datasets focused on streaming data, that would be really helpful - we could > add them to the mix. > > In the meantime I can make some high level observations: > > 1) HTM's are not more computationally efficient. There's a lot more going on. > It is reasonably fast - it takes about 25 msecs on my laptop to process one > data point. MA models are probably an order of magnitude faster though. > > 2) HTM's can learn many different patterns, not just one pattern. It can also > make multiple predictions at each time step. With the likelihood code it can > also learn high level statistics about the data and thus can work with highly > unpredictable data. > > 3) Anecdotally, with Grok, we've seen many examples of HTM anomaly detection > that probably would not have been caught by MA models. > > 4) I did do a comparison of HTM prediction against ARIMA (I believe ARIMA is > a superset of Holt-Winters). HTM did a lot better on one energy dataset. > There was a discussion of it here: > > http://lists.numenta.org/pipermail/nupic_lists.numenta.org/2013-September/001411.html > > Hope this helps! > > --Subutai > > > >> On Fri, Jul 25, 2014 at 10:22 AM, Omair Shafi <[email protected]> wrote: >> Hi again, >> >> I've been fooling around the Anomaly detection sample code by Subutai using >> TemporalAnomaly, which seems absolutely awesome! I just wanted to know how >> it compared with standard models for anomaly detection like Holt Winters, >> other variations of Moving Average and the lot. Does it score higher on >> accuracy? Or is it more computationally efficient? Or maybe it can learn >> more than one pattern? >> >> Thanks! >> >> Regards, >> Omair S >> >> _______________________________________________ >> 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
