On 02/15/2016 05:19 PM, Matthew Taylor wrote:
See below...
On Sat, Feb 13, 2016 at 6:59 PM, Wakan Tanka <[email protected]> wrote:
Thank you very much Matt, informative as always ;)
One more questions:
2) Swarming is optimized only for prediction. It may not be the best
method to find model params for anomalies. We have identified a set of
model params that are decent for most one-dimensional scalar input
anomaly detection, and we generally reuse those in all our anomaly models.
So is there better way to find model params for anomaly detection than
swarming?
1. Isn't anomaly just prediction where NuPIC missed. Why is such difference
between anomaly and prediction during swarming?
Swarming is only for prediction. The process specifically uses
prediction accuracy as a goal when permuting over model params. It
completely ignores anomalies. It is impossible to rate how well
anomaly detection is doing without labeled input data that calls out
where the actual anomalies are. Swarming cannot do that.
OK, maybe some video tutorial for advanced users would be nice ;)
2. Is it possible to somehow display and being able to read something useful
from the internal state of model object?
Probably not unless you know an awful lot about the theory, and even
then, our engineers only seem to need to look into the cellular state
when they are debugging something unexpected.
I know that with using TemporalAnomaly I can predict one step as I were
choose TemporalMultiStep, AFAIK the TemporalMultiStep allows me to
predict multiple steps at once and TemporalAnomaly allows me to predict
one step (with choosen size) plus anomaly score. I'm just curious why
people from Numenta decided to separate them, is it only for
peroformance purposes? Also if I understand correct then predicting
larger step will have significant impact on performance but predicting
more will not. I guess that when NuPIC predicts 10 steps ahead then it
under the hood also predicts steps from 1 to 9. So it is more a matter
of memory than CPU. Or am I wrong?
3. Regarding inferenceType: is there any type which is combination of
TemporalMultiStep and TemporalAnomaly or is there any reason why NuPIC cannot
output multiple steps and also anomalies at the same time?
If you use TemporalAnomaly, you'll still get predictions out. So if
you want both you can have both, it just increases the processing
time.
Regards,
---------
Matt Taylor
OS Community Flag-Bearer
Numenta
