"Keep an eye on this space for information about the upcoming open sourcing of the Numenta Platform for Intelligent Computing, which includes the the Cortical Learning Algorithm used by Numenta to create our flagship product, Grok." (Numenta.org)
Something in my mind says this is related in some way to the start/progress of our community.. he he
On the one hand, it's good because we will have updated version of Grok and several people will can develop applications with the best of the HTM, on the other hand, Numenta could stop any time the support to the community or then leave very limited the participation of the community members (it's just see past experience)..
Anyway, thanks for the news, and let's "keep an eye" there.
UPDATE: I'm thinking this Numenta.org very, very weird.... All main links direct to numenta.org, not numenta.com. No news related on official Numenta site.. I hope I'm wrong.
Last edit: David Ragazzi 2013-04-26
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
The github site links to numenta.com. And the account exists since 2011, so Numenta would probably have objected if this is not genuine. (It also contains several language bindings for the Grok api, fakeing those would really be - weird).
Interestingly the new site also mentions a (new?) book "On Biological Computing" by Jeff Hawkins...
But time will tell...
Imho OpenHTM will be helpful in any case. The gui could be used to visualize different CLA implementations. And the theory offers such a broad field of interpretations and goals to optimize for that there is definitely room for more than one implementation. And implementing something yourself leads to a higher level of understanding than just reading about it. And there might be some developers who don't like to contribute under a contributers agreement that reserves all (e.g. commercial) rights to one company. So keep on the good work!
But it's give and take - if Numenta releases sufficient amounts of code so the open source release is useful in itself and they search the dialogue with the community then I wouldn't have objections to contribute back - everybody can only win then.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Interesting news - thanks for point this out. I think it's legitimate. The GitHub page is maintained by Matthew Taylor who apparently works at Numenta.
My wild guess is this will not be an open version of Grok, but some sort of web shell, or frontend that will use the Grok APIs to access Grok running on Numenta's server. I don't think they will open up the Grok code.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hey Ferdinand thanks for bringing this to our attention! This looks like it could be very promising. I don't see any algorithm source code in the github so far, but the site does imply that it may be coming soon. I sure hope it does as it would be very interesting to see how Grok's core really works. So many questions about the original CLA document that came out 2 years ago. Seeing the current Grok source would hopefully bring us up to date in short order.
It also appears we may get more clarification on the licensing terms for CLA, which is something else we have been a bit fuzzy on so far. Perhaps some of our work here can be contributed back to Numenta at some point, yet another interesting possibility to consider. I know I would not mind a more open conversion with the Numenta team about our work on OpenHTM.
I will certainly be keeping an eye on numenta.org in the coming days to see what develops.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
As far as I understood from new "CLA Basics" vids:
They use 4 main classes:
-sensor/encoder
-SP
-TP
-CLA classifier
Encoder converts the categorial/numerical input data into a new, semantic, binary represntation of variable length.
they use 1-dim vector of columns in a region.
SP has 2048 columns, 40 active.
TP has 32 CPC. It's output is 32x40 = 1280(?) active only bits. (He says 65536, why?)
Classifier makes single step predictions but for t+K steps, it associates every cell with a lookup table with an element of input 'bucket' index from the Sensor and its corresponding active steps.
The guy also talks about 'scaler encoders' with 21 ones and optimized overall size, which are used for converting numerical ranges to semantic representations. He also mentions 'moving average' being involved in prediction.
Edit: they DO use each cell output.
Last edit: Nick 2013-05-28
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Something seems to be going on at Numenta...
https://github.com/numenta/nupic/tree/gh-pages
Looks like the core engine of Grok goes GPL - interesting! :)
http://numenta.org/licenses/
edit: can an admin please move this post to the General Discussions sub-forum? Thanks!
Last edit: Ferdinand Strixner 2013-04-26
"Keep an eye on this space for information about the upcoming open sourcing of the Numenta Platform for Intelligent Computing, which includes the the Cortical Learning Algorithm used by Numenta to create our flagship product, Grok." (Numenta.org)
Something in my mind says this is related in some way to the start/progress of our community.. he he
On the one hand, it's good because we will have updated version of Grok and several people will can develop applications with the best of the HTM, on the other hand, Numenta could stop any time the support to the community or then leave very limited the participation of the community members (it's just see past experience)..
Anyway, thanks for the news, and let's "keep an eye" there.
UPDATE: I'm thinking this Numenta.org very, very weird.... All main links direct to numenta.org, not numenta.com. No news related on official Numenta site.. I hope I'm wrong.
Last edit: David Ragazzi 2013-04-26
The github site links to numenta.com. And the account exists since 2011, so Numenta would probably have objected if this is not genuine. (It also contains several language bindings for the Grok api, fakeing those would really be - weird).
Interestingly the new site also mentions a (new?) book "On Biological Computing" by Jeff Hawkins...
But time will tell...
Imho OpenHTM will be helpful in any case. The gui could be used to visualize different CLA implementations. And the theory offers such a broad field of interpretations and goals to optimize for that there is definitely room for more than one implementation. And implementing something yourself leads to a higher level of understanding than just reading about it. And there might be some developers who don't like to contribute under a contributers agreement that reserves all (e.g. commercial) rights to one company. So keep on the good work!
But it's give and take - if Numenta releases sufficient amounts of code so the open source release is useful in itself and they search the dialogue with the community then I wouldn't have objections to contribute back - everybody can only win then.
Interesting news - thanks for point this out. I think it's legitimate. The GitHub page is maintained by Matthew Taylor who apparently works at Numenta.
My wild guess is this will not be an open version of Grok, but some sort of web shell, or frontend that will use the Grok APIs to access Grok running on Numenta's server. I don't think they will open up the Grok code.
Hey Ferdinand thanks for bringing this to our attention! This looks like it could be very promising. I don't see any algorithm source code in the github so far, but the site does imply that it may be coming soon. I sure hope it does as it would be very interesting to see how Grok's core really works. So many questions about the original CLA document that came out 2 years ago. Seeing the current Grok source would hopefully bring us up to date in short order.
It also appears we may get more clarification on the licensing terms for CLA, which is something else we have been a bit fuzzy on so far. Perhaps some of our work here can be contributed back to Numenta at some point, yet another interesting possibility to consider. I know I would not mind a more open conversion with the Numenta team about our work on OpenHTM.
I will certainly be keeping an eye on numenta.org in the coming days to see what develops.
As far as I understood from new "CLA Basics" vids:
They use 4 main classes:
-sensor/encoder
-SP
-TP
-CLA classifier
Encoder converts the categorial/numerical input data into a new, semantic, binary represntation of variable length.
they use 1-dim vector of columns in a region.
SP has 2048 columns, 40 active.
TP has 32 CPC. It's output is 32x40 = 1280(?) active only bits. (He says 65536, why?)
Classifier makes single step predictions but for t+K steps, it associates every cell with a lookup table with an element of input 'bucket' index from the Sensor and its corresponding active steps.
The guy also talks about 'scaler encoders' with 21 ones and optimized overall size, which are used for converting numerical ranges to semantic representations. He also mentions 'moving average' being involved in prediction.
Edit: they DO use each cell output.
Last edit: Nick 2013-05-28
Strange: Hawkins writes they have 10 cells per column
Last edit: Nick 2013-05-27