Hi,

On Sun, Apr 27, 2014 at 1:33 AM, Ritchie Lee <[email protected]> wrote:

> Hi friends of NuPIC,
>
> Here is another example that I ran in my quest to understand how CLA
> works.  This is a 1D example where the signal is a discrete step from 1 to
> 2 with small random Gaussian noise added throughout. Learning is on for the
> first 5000 and off for the remainder.  i.e., CLA training does not see the
> anomaly.  Here are the plots:  1. signal in black, predicted in blue.  2.
> residual = abs(predicted-signal) 3. anomaly score.
>


> I specified specifically that CLA should use a ScalarEncoder with min=0.8
> and max=2.2.
>
just a check, did you use "enough bits" in `w` for good results and
`resolution`, right?


>  Now the question is why does the anomaly score not react to the anomaly
> at all?  I expect it to spike at the step and in addition to that stay
> excited throughout the unseen portion.
>
Great! this would confirm my theory in the previous mail to you. (#3) is a
bug! maybe we can return last-seen as predicted (active cells), but we must
then set predictive cells to all 1s (anomaly)


-- 
Marek Otahal :o)
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to