> > From your post it appears that you are getting around a lot of the issues > with implementing the temporal pooler on GPU by essentially just having a > single distal segment per cell, which has a synapse (a weight) for each > cell in its input area, rather than adding synapses over time as needed?
Yes, it starts out with all the synapses it will ever need. Growing new connections on the GPU would be rather difficult (although doable). Do you find that this solution gives you good quality of predictions, even > in cases where a pattern may be followed by a large number of subsequent > patterns? I don't know how good the predictions really are yet, I just know that it works at all. I haven't done any testing beyond just moving a few patterns around the screen and predicting them one step ahead of time, which worked. Also, what kind of speedup are you seeing for the GPU version vs the CPU > version? I haven't measured it for my original discrete HTM, but I get 80 steps per second for a region with 16384 columns, 65536 cells, 1.3 million column connections and 21 million lateral connections. I haven't really optimized it yet though. On Tue, Oct 21, 2014 at 5:25 PM, Michael Ferrier <[email protected]> wrote: > Hi Eric, > > This is interesting to me, because I'm also working on a continuous > version of HTM that runs on the GPU. From your post it appears that you are > getting around a lot of the issues with implementing the temporal pooler on > GPU by essentially just having a single distal segment per cell, which has > a synapse (a weight) for each cell in its input area, rather than adding > synapses over time as needed? Do you find that this solution gives you good > quality of predictions, even in cases where a pattern may be followed by a > large number of subsequent patterns? Also, what kind of speedup are you > seeing for the GPU version vs the CPU version? > > Thanks! > > Mike > > _____________ > Michael Ferrier > Department of Cognitive, Linguistic and Psychological Sciences, Brown > University > [email protected] > > On Mon, Oct 20, 2014 at 4:27 PM, Eric Laukien <[email protected]> > wrote: > >> Hello everyone, >> >> I'm new to this community. I have recently taken interest in HTM. I come >> from the reinforcement learning scene, so naturally I want to use HTM to >> perform reinforcement learning. I have designed a new HTM-like algorithm >> for this task that I would like to share here. >> >> My version of HTM differs mainly from the original in that it operates on >> continuous states. A cell isn't just on or off, neither is a column, it can >> also be in-between. The use of continuous states sounds like it would make >> the algorithm more complicated, but actually it has done the opposite. >> Continuous HTM, at least in the form I have it now, is much less code than >> the original (which I also implemented). >> >> Aside from being smaller, continuous HTM has some other benefits: It runs >> really well on the GPU (I have it running nicely on the GPU right now), and >> the continuous states allow you to encode more information in fewer columns. >> >> I don't know how biologically plausible this is, but it may or may not be >> more practical than standard HTM. >> >> I have a blog where I describe my attempts at using HTM to outperform >> Deepmind's Atari results. In this blog I describe in detail the operation >> of continuous HTM. Link: http://cireneikual.wordpress.com/ >> >> I would like to get your thoughts on this algorithm. I will be releasing >> the source code soon, so you can experiment with it yourself. >> >> I wonder how difficult it would be to modify NuPIC to incorporate >> continuous states. >> >> Thank you for your attention! >> > >
