I just had a thought, and its do with the hierarchy, and inference, and spacial
pooling.
when you are inferring, you wind up, every symantic, a percentage error, coming
up the proximals,you can use a threshold to instantly activate it at some
error. but wouldnt it be better, if you just passed up the error? it works
better this way, because then your smaller groups will work better, because
they dont get stuck at an error minima.the later symantics can then
"continuify" the error to a more exact quotient, by reading the error of the
previous level, usingthe errors together to make a new error, which would have
less error than the previous level, most of the time.
then you could keep passing the error till the last region, then have an exact
"nearness" to the novel inputusing every single symantic together, leaving
activation till later, making use of the distribution together.
notes about my system->Its purely spacial, with no temporal yet, that makes
this idea much easier to do. i also finish at a singlecollection, so all i
have to do to inference is pick the cell of least error. and it also has a
separate inferencing stage from trainingso i dont know if this idea applies to
htm, but possibly could be useful to know.
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
