Closed rhyolight closed 10 years ago
subutai
commented
10 years ago We have tried variants of this before and after Rahul. We ran into a bunch of issues since the SDR changes significantly over time. Closing this issue.
Note: Jeff has an idea to do this using the random encoders, which would avoid changing the SDR. It's low priority though.
fergalbyrne
commented
10 years ago Rahul,
We had a pretty extensive discussion about this on the mailing list last Summer. There is, in fact a frequency-based adaptive encoder [1] but I'm not sure if it's fully functional.
F
[1] https://github.com/numenta/nupic/blob/master/nupic/encoders/nonuniformscalar.py
On Mon, Aug 4, 2014 at 5:49 PM, Subutai Ahmad notifications@github.com wrote:
We have tried variants of this before and after Rahul. We ran into a bunch of issues since the SDR changes significantly over time. Closing this issue.
Note: Jeff has an idea to do this using the random encoders, which would avoid changing the SDR. It's low priority though.
— Reply to this email directly or view it on GitHub https://github.com/numenta/nupic/issues/1162#issuecomment-51085642.
Fergal Byrne, Brenter IT
Author, Real Machine Intelligence with Clortex and NuPIC https://leanpub.com/realsmartmachines
Speaking on Clortex and HTM/CLA at euroClojure Krakow, June 2014: http://euroclojure.com/2014/ and at LambdaJam Chicago, July 2014: http://www.lambdajam.com
http://inbits.com - Better Living through Thoughtful Technology http://ie.linkedin.com/in/fergbyrne/ - https://github.com/fergalbyrne
e:fergalbyrnedublin@gmail.com t:+353 83 4214179 Join the quest for Machine Intelligence at http://numenta.org Formerly of Adnet editor@adnet.ie http://www.adnet.ie
From: https://issues.numenta.org/browse/NPC-170
I (Rahul) have an idea for a new adaptive encoder that is non-linear, allocating buckets based on the frequency distribution of the data. This could be especially useful for adding learning to the encoder->TP models, which we are creating for anomaly detection.
The idea is that each bit in the encoder has a certain "receptive field" in the input space, and it can expand and contract based on how frequently it fires. It should also have the property that, for very simple input spaces, there is not oscillation.