Magnus, That's right. Every dendrite segment has some number of synapses and a threshold. If the number of synapses formed to recognize a pattern is 20 and the threshold is 15 then you can have up to 5 missing inputs before the segment fails to recognize the pattern. Other segments will typically have low overlap with the pattern so they would need 10 or more extra noise input bits before falsely recognizing something. It is hard to say exactly how much noise will be ignored but the basic idea applies to both the spatial pooling synapses and the sequence memory synapses.
This is already part of the CLA. Jeff From: nupic [mailto:[email protected]] On Behalf Of Magnus McWootton Sent: Thursday, February 20, 2014 8:54 AM To: [email protected] Subject: [nupic-discuss] distributed representations I was just thinking a little about SDRS... A distributed representation has many more states possible, by reusing cells, as aposed to a nondistributed rep (which is 1 cell) and then it has as many different states as cells only. But I was wondering, can the distributed reps in HTM can actually clean up an errored rep to one its seen before? (using the distil dendrites!) I figure it would be a matter of teamwork between the cells to get to a whole distribution again. Would be good for completing noisy or missing images to a closest guess... and being distributed somewhat youd benefit from being distributed and not just 1 cell per column, and have more total states. I was thinking if it worked you could do it every level in the hierarchy, as a "stabalizer" or something.
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
