David, I'd also like to mention that I'm beginning work on a simplified temporal memory (what you called sequence memory) implementation, that's closer to the white paper description. This work should be completed in the near future. However, you can still definitely port the existing temporal memory (currently called TP) implementation to Java if you don't want to wait for the simplified one.
- Chetan On July 14, 2014 at 9:02:58 AM, Yuwei Cui ([email protected]) wrote: Hi David, Each layer will be implementing sequence memory (we now call it temporal memory). The goal is to predict future input based on past inputs. Spatial pooling and sequence memory will not be combined. If you have read the white paper, the spatial pooling occurs on the proximal dendrite, which can drive cells into "active states", and sequence memory relies on inputs on distal dendrite, which will turn cells into "predicted states" Our current working hypothesis is temporal pooling occurs both between layer 4 and layer 3, and across regions. The outcome of temporal pooling is more stable representation higher in the hierarchy. There will be an emphasis on the stability soon, we already have a working temporal pooling algorithm and will release it in the near future. The inter-regional feedback seems a little further down the road. Yuwei On Sun, Jul 13, 2014 at 3:12 PM, cogmission1 . <[email protected]> wrote: Hey Matt, I just watched the "Fireside chat" about Jeff's next TP ideas, and not to be argumentative (just looking to be as clear as I can be while watching things happen from the outside), but it appears as if he's going to introduce sequence memory, and temporal pooling between each layer? I think definitely between within layers 3 and 4? Also, it appears as if the SP and the Sequence Memory will be combined? Additionally, it looks like there will be some "inter-regional" feedback/predictions introduced between a subsequent region's layer 6 back into layer 5 of the previous layer. It maybe looks like some drastic overall changes may be occurring - maybe due to an effort toward emphasizing concepts of stability and inter-regional feedback?. Any news as to when we could see these changes rolled out - I can hardly wait...? :-P On Sat, Jul 12, 2014 at 12:08 PM, Matthew Taylor <[email protected]> wrote: You can definitely dive in now. The new temporal memory work isn't going to disrupt the current architecture at all, as far as I know. Start here: https://github.com/numenta/nupic/wiki/Learning-NuPIC --------- Matt Taylor OS Community Flag-Bearer Numenta On Sat, Jul 12, 2014 at 9:42 AM, cogmission1 . <[email protected]> wrote: Hi Everyone, I'm at the point now where I would begin to delve into the NuPIC code more deeply (I want to port it to Java as a way to exercise my understanding too), but I was wondering if there is any point to it prior to the new TP/Sequence Memory implementation? Should I maybe wait? Thanks David _______________________________________________ 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 -- -- Yuwei Cui Algorithm Internship, Numenta Inc. PhD Candidate, Neuroscience and Cognitive Science University of Maryland, College Park, MD, 20742 Homepage: http://terpconnect.umd.edu/~ywcui/ _______________________________________________ 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
