There are already a couple of teams working on HW implementations of the
CLA.  IBM Research and Seagate are working on pure CLA HW.  There is a
scientist at Sandia National Labs also interested.  And there is a program
director at DARPA putting together a HW initiative for temporal learning
algorithms within hierarchies.  That was inspired by the CLA.  These
initiatives are very early but the principle scientists involved are serious
about this.  At some point it might make sense to have an entire section for
people working on HW implementations.  If anyone wants to be introduced to
these scientists let me know and I will query their interest.

Jeff

 

From: nupic [mailto:[email protected]] On Behalf Of Fergal
Byrne
Sent: Tuesday, July 16, 2013 2:29 PM
To: NuPIC general mailing list.
Subject: Re: [nupic-dev] Newbie question

 

 

Hi Subutai,

 

Brilliant to hear the historical background, don't misunderestimate (as GW
Bush said) the importance of that for understanding of how the CLA works.

 

One of Jeff's big motivations for the work you guys have been doing is that
we (as humans) are able to do these amazing things with a very limited
number of really slow neurons, with no ability to "upgrade" or, for example
add hard drives for extra storage.

 

We just have to be using a very clever, but very easily implemented,
algorithm for this performance which is robust in every dimension of energy,
material and time.

 

So, a CLA (a 1mm squared, 1-layer, one region slice) should be implementable
in any kind of computational environment, in any language, to an order of
magnitude of performance. 

 

Developer and experimenter time is the limiting factor in the work right
now. Once we prove this is how brains work, someone will (and probably is
already working to) build this in hardware or figure out how to get your
graphics card to process this.

 

Regards,

 

Fergal Byrne

 

 

_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to