Fergal,
I look forwards to the the new Temporal Pooling (TP) implementation and
hopefully, by then,
we'll be able to use some hierarchy as well. As for GPU version of
Nupic, I just meant that
unless we could implement Nupic (with the new TP and hierarchy) to take
advantage of GPU platform
such as CUDA or other parallel computing platforms, it would probably be
very difficult to train
Nupic with videos within reasonable timeframe. After all, one of the
main factors behind deep
learning success is that they can scale up their networks with NVIDIA
GPU cards and/or CPU
clusters (e.g., Google, Microsoft).
Regards,
Vinh
On Tuesday 30,September,2014 01:24 AM, Fergal Byrne wrote:
Cheers, Subutai.
Vinh,
On NLP: cortical.io <http://cortical.io> and NuPIC are already able to
do some really interesting things, but I believe this will really open
up when we combine cortical.io <http://cortical.io> on real textual
streams with the L4-L3 Temporal Pooling and hierarchy.
On Video: the Deep Learning guys have made some pretty dramatic
progress on this recently - they seem to be using a type of
multi-resolution rolled-out hierarchy and a sort of saccading viewport.
This is a great example of how the slow brain deals with a million
channels of fast-changing input data streams. The neocortex learns
which information to throw away and still keep so much structure that
it can do vision which dedicated computer systems cannot match. It
seems likely to me that we won't require hardware to achieve similar
feats with HTM; it can be done with a software design built from the
outset with parallelism and concurrency, and aggressively exploiting
sparsity (perhaps like the one I'm working on ;).
Regards,
Fergal Byrne
On Mon, Sep 29, 2014 at 5:18 PM, Subutai Ahmad <[email protected]
<mailto:[email protected]>> wrote:
On Mon, Sep 29, 2014 at 6:31 AM, Fergal Byrne
<[email protected] <mailto:[email protected]>>
wrote:
What they say, and perhaps it's worth thinking about, is that
we would gain credibility if we could demonstrate a "killer
application" (what Ben Goertzel calls the "AGI Sputnik
moment") of HTM which shows it solving a problem nobody else
can even attempt to solve. For that, I believe, we'll need the
full sensorimotor stack and hierarchy, which we will have in
the next few months.
Could I ask people to have a think about this and possibly
bounce around ideas? We could schedule a round-table session
during the hackathon next month and see if there's an
application area to focus on in this regard. The most
immediate candidates I can see right now are cortical.io
<http://cortical.io> (aka CEPT) for NLP and the Geospatial
Encoder.
Hi Fergal,
This is a really nice idea. I'm sure Jeff will be happy to
participate as well. Perhaps it is a discussion we can carry on
during office hours as well.
--Subutai
--
Fergal Byrne, Brenter IT
http://inbits.com <http://inbits.com/> - Better Living through
Thoughtful Technology
http://ie.linkedin.com/in/fergbyrne/ - https://github.com/fergalbyrne
Founder of Clortex: HTM in Clojure -
https://github.com/nupic-community/clortex
Author, Real Machine Intelligence with Clortex and NuPIC
Read for free or buy the book at 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
e:[email protected]
<mailto:[email protected]> t:+353 83 4214179
Join the quest for Machine Intelligence at http://numenta.org
Formerly of Adnet [email protected] <mailto:[email protected]>
http://www.adnet.ie <http://www.adnet.ie/>
