Hi Dionysis,

The Geospatial encoder is designed to produce similar SDRs for similar inputs. 
Adding a bit of noise to an input should still produce a similar SDR, so the 
sequence learning algorithm that the anomaly detection is based on should still 
be able to recognize the inputs. See the geospatial demo [1] for an example of 
this (no two routes are exactly the same, but it still learns them fine).

[1] https://github.com/numenta/nupic.geospatial 
<https://github.com/numenta/nupic.geospatial>

- Chetan

> On Dec 16, 2015, at 1:35 AM, Dionysis Manousakas <[email protected]> wrote:
> 
> Hi NuPIC,
> 
> I'm trying to test the NuPIC Geospatial encoder on noisy data.
> So, I traversing a whole period of sinusoidal trajectory with no noise
> (you can think of angle as longitude and sinus as latitude) 
> and then repeat on noisy versions of the signal. The anomaly detector 
> tends to entirely flag the noisy trajectories as anomalies. Is this the
> result to be expected or should the predictions be robust to noise?
> 
> Best regards,
> Dionysis 

Reply via email to