Thank you Hannu, I chose OpenCL for its platform independence (and simplicity) even though it doesn't come with any ready-made solutions for linear algebra etc. After all the first phase of the spatial pooler is basically sparse matrix multiplication. The neighbourhood inhibition step is much harder to parallelize..
My motivation? What could be more interesting!? :) Initially I was struck by how similar this is to the methods used in "traditional" & mathematics-backed machine learning and how nicely it incorporates methodology like Hebbian learning. My knowledge in neuroscience is limited to what I read in Jeff's book, but he really knows how to make the field more approachable to a software engineer. As Matthew added the IRC channel is located at irc.freenode.net. Joonas Haapala 2013/10/25 Hannu Kettinen <[email protected]> > Hi Joonas, > > Impressive work. > > What IRC Server are you on? (Haven’t used IRC since the 90´s :) ) I have > a pending request for clearance to share some CUDA code that might interest > you. Very similar to openCL. > > I do not know your purpose for working on this, but I see you like to work > in an linux environment. One piece of advice I can give you is to read up > on powerLinux a bit: > http://www-03.ibm.com/systems/power/solutions/developers/ > This is the hardware platform where the revolution will take place, and > IBM’s watson runs on power systems. (And no I do not work for IBM). Today > we can have 16TB of RAM on the 795 and that will soon double to 32TB > according to some rumours regarding the power8 platform. > > I am sorry I can only offer my encouragement at this point in time, but > hopefully can soon share some code that might interest you. > > > > Good luck and keep up the good work! > > -Hannu > > > > > > > > > > On 25 Oct 2013, at 00:10, Joonas Haapala <[email protected]> wrote: > > > Hello Numenta! > > I'll start with a quick introduction: I'm a student at Aalto University, > Finland, currently studying for B.Sc. in Systems Analysis (applied maths, > statistics, decision theory and the like). > > I remember seeing a discussion about porting the HTM algorithms to a GPU > for performance boosts (and someone mentioning on the IRC channel that it > will be hard, and I'm not here to prove otherwise) so I decided to give it > a shot. > > In my free time I've been working on an OpenCL implementation of the > Cortical Learning Algorithms as described in the whitepaper. The project is > still in very early stages and I haven't any cool results to show, but I'm > hoping it will serve at least as a source of inspiration to the community. > > I try to keep the code as simple as possible so that it is still possible > to reason about the logic behind it. That being said I had to cut some > corners in the process of porting the algorithm due to the limitations of > SIMT computing and I haven't documented all those changes yet. One such > change is that in the TP I never build a list of proposed changes to a > segment, I just collapse the changes to a separate state and later decide > whether to apply them or not. > > The code is available in GitHub: http://github.com/Jontte/CortiCL > > I'm a keen IRC user so should you prefer a more informal chat with me you > will find me in the #nupic channel with the nickname Jontte. > > Best regards, > > Joonas Haapala > _______________________________________________ > 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
