All of these are great suggestions, but I would have to say that Karthik's #2 suggestion resonates most strongly for me. A demonstration of NuPIC with a compelling application. A use case that showcases in certain terms what NuPIC can do that no other approach can do.
- Jeff On Mon, Jan 6, 2014 at 6:49 PM, karthikeyan S <[email protected]>wrote: > Hi > > Just my 2 cents. > > 1. As others mentioned, complete c++ implementation > 2. Demonstration of NuPIC with a compelling application by the community. > I think instead of competing with other ML techniques we should try and > exploit what's unique about NuPIC - Which, I think, is mimicking human > thought process. Other ML techniques are fairly problem specific and we > have a slight disadvantage competing with them. And more importantly there > is not much point in competing since there are already tools to solve > *those* specific problems. > We don't want to be an yet-another. > An example of what I mean by mimicking human thought process is math and > physics. > > Math - if we "teach" NuPIC arithmetic will it learn? > Can it add 2+2 without an ALU. > Physics - If NuPIC constantly gets input about its environment through its > sensors will it form any laws or theories from it, at some point. The > current anomaly detection (something didn't comply with nupic's internal > theory of its environment)is close to this. But can we stretch it? > > As they say, sometimes it's easier to do the hard things and it's more fun. > > Just my 2 cents. > > Thanks > Karthik > On Jan 6, 2014 6:03 PM, "stewart mackenzie" <[email protected]> wrote: > >> - a core c++ implementation stripped of network manager, that consists >> of encoders, regions and classifiers. (the community and grok can >> extend this repo with customized regions etc forming a repo with a >> whole bunch of different tools) >> --- python completely stripped out >> --- nupic-core should have no library dependencies. >> --- temporal pooler implemented in c++ >> - each component (encoder, region classifier) knows how to read the >> data stream coming into it. whether it be 1D or 2D etc (hopefully >> using UBF - by J.Armstrong) >> - CMake support. >> - multi-time interval predictions (ie what the next 5 mins predict? >> then what does the next 20 hours predict?) >> - an implementation that is _easy_ to create language bindings for. >> >> _______________________________________________ >> nupic mailing list >> [email protected] >> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org >> > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > >
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
