thanks, Yes the connection is one-to-one in one sense. /If you pass randomize=X.XX parameter it will create some random connections, but that is beside the point./
And here is where my main speculation with this prj is. One abstraction as I explained is the cell-to-cell full connectivity, (by default there is no connection), but there is another abstraction. Once the system start learning all the possible connections are regarded as union-SDR, so I'm speculating that union-SDR behaves discriminatory as a bunch of separate groups-of-dendrites, because it acts only on specific patterns and not others (by means of overlap), the same as organically growing dendrites will behave i.e. detect patterns. At least that is the hope. But you caught rightly the main contention. I don't have the dying part yet, but have couple of ideas. My current tests don't fill the memory at all, so it is save for now. -------| http://ifni.co On Thu, May 5, 2016 at 7:28 AM, cogmission (David Ray) < [email protected]> wrote: > Hi, > > Very nice read! I enjoyed following you as you went through the steps in > learning the finer points of the theory. It was nice to compare with my own > notions and confusions regarding the different concepts. I also enjoyed the > section on NN's which I have no previous history with, and have been > hesitant to immerse myself in learning since it isn't the way to the > future. It's hard to go back and learn something for the benefit of > perspective when the status quo has already moved on to something > different. Weird predicament. I almost would have rather learned NN's > before hearing about HTM Theory - but alas that is not the situation. > > Observation: > Here is a small critique (since you asked for them specifically) ;) > > One possible critique is that you seem to model 1-to-1 connections between > cells (dendrites)? As far as I have seen from Jeff's discussions is that > the primary learning method is the _growth_ of dendrites (i.e. a single > cell can connect to many different cells via its dendrites) - so there may > be a problem with the "online learning" portion of its functionality? In > other words the use of the same system for one or more different problem > domains without turning it off? There may be other implications though - > which I don't have time to write about right now. > > Anyway, I enjoyed the read and look forward to following your journey as > you improve/optimize your system! Thank you for sharing your work! > > Cheers, > David > > On Thu, May 5, 2016 at 4:20 AM, Wakan Tanka <[email protected]> wrote: > >> Very nice, thank you >> >> On Thu, May 5, 2016 at 4:23 AM, mraptor <[email protected]> wrote: >> >>> hi guys, >>> >>> Thanks for the HTM. >>> I made standalone implementation of HTM as I understand it at the moment >>> written entirely in Python. >>> In the following project I tried to consolidate both documentation and >>> code in single package. >>> >>> Here is the code : >>> https://github.com/vsraptor/bbhtm >>> and here are the docs : >>> http://ifni.co/bbHTM.html >>> >>> I hope you like it :) >>> Keep in mind it is still very minimal implementation primary of TM for >>> now and need alot more testing, but I could not wait to share it, so I can >>> confirm I'm more or less on the right path, before I try more crazy stuff >>> :). >>> >>> I would love your feedback. >>> >>> -------| http://ifni.co >>> >> >> >> >> -- >> Best Regards >> >> Name: Wakan Tanka a.k.a. Wakatana a.k.a. MackoP00h >> Location: Europe >> Note: I'm non native English speaker so please bare with me ;) >> Contact: >> [email protected] >> http://stackoverflow.com/users/1616488/wakan-tanka >> https://github.com/wakatana >> https://twitter.com/MackoP00h >> > > > > -- > *With kind regards,* > > David Ray > Java Solutions Architect > > *Cortical.io <http://cortical.io/>* > Sponsor of: HTM.java <https://github.com/numenta/htm.java> > > [email protected] > http://cortical.io >
