Hi

 I was studying the TP algorithm and didn't understand one point.

 In the beginning, permanence values are assigned aleatory to distal
dendrites segments of all cells (as in the initialization of the SP).

 In time t, the region predicts a bunch of cells based on the activation of
their segments. A segment is active if it have a minimum number of active
synapses connected with active cells in time t. In time t + 1, the cells
that were predicted and are in active columns (and just that in active
columns) will have their synapses permanence values updated according.

 But, what about the false positives, the cells that were predicted
ocasionaly just because they had an active segment that fit well with input
in time t, but they never occur. Don't they will be predicted again and
again when the input in time t occur? I couldn't see where the algorithm
treat this issue.

 In the last part, it does

for c, i in cells
55. if learnState(c, i, t) == 1 then                                =>
update just the cells that are in an active column
56. adaptSegments (segmentUpdateList(c, i), true)
57. segmentUpdateList(c, i).delete()

Doesn't necessary to decrement the permanence value of those connected
synapses in the distal dentrite segments of cells that  don't occur?


Christian
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to