Manal, Since we were aiming to make a demo geospatial application, my main goal was to have something that caught relevant anomalies and learned fast. In deciding on parameters, I visually inspected the results for a couple of routes that I had generated by tracking my commutes from work to home. Then I hand-tuned parameters to get better visual results on these routes.
For a real application, we would probably use a larger dataset and use swarming with the error metric being accuracy on hand-labeled anomalous regions of the routes. But we have not done this yet. - Chetan > On Nov 19, 2014, at 12:43 PM, Manal Bhingardeve <[email protected]> > wrote: > > Hi, > So based on these changes what error metric was used for evaluating the > effectiveness of the model and how was this model found to be better? Were > different permutations run programmatically? Which data set was used to > evaluate the performance , is it publicly available? > Thanks, > Manal > > On Wed, Nov 19, 2014 at 11:57 PM, Chetan Surpur <[email protected] > <mailto:[email protected]>> wrote: > Hi Manal, > > I started with the model params for hotgym, and updated a couple parameters > for better performance. > > In particular, > > 1. Removed the timestamp parameter (leaving only the geospatial encoder's > output as a parameter) > 2. Matched encoder width to number of columns (2048, same as hotgym example), > keeping sparsity the same > 3. Updated parameters: minThreshold 9 => 3, activationThreshold 12 => 6 (for > faster prediction) > > - Chetan > > >> On Nov 19, 2014, at 5:21 AM, Manal Bhingardeve <[email protected] >> <mailto:[email protected]>> wrote: >> >> Hi, >> Could you share more details about the process that was followed to derive >> the parameters in the >> model_params.py (Geospatial Tracking application) file. Is there a generic >> process that can be used for prediction on multiple fields? >> Thanks, >> Manal >> >> On Thu, Nov 13, 2014 at 10:41 PM, Matthew Taylor <[email protected] >> <mailto:[email protected]>> wrote: >> The swarming process optimizes for prediction on one specific field in >> the input data. So it does not always produce model params that are >> best for anomaly detection. Chetan used some hand-created model params >> for that example application. >> --------- >> Matt Taylor >> OS Community Flag-Bearer >> Numenta >> >> >> On Thu, Nov 13, 2014 at 9:07 AM, Manal Bhingardeve >> <[email protected] <mailto:[email protected]>> wrote: >> > Hi, >> > >> > In the NuPIC Geospatial Tracking application example how is the >> > model_params.py file obtained? Shouldn't there be a swarm file? >> > >> > Thanks, >> > Manal >> > >> >> > >
