Ensure new temporal memory contains learn on one cell bug fix about htmresearch HOT 5 CLOSED

@subutai When you say "bug fix", what bug are you referring to?

from htmresearch.

Sorry I left out the actual bug fix in that snippet! I think the previous version of the code did not go through self.getBestMatchingSegment if self.learnOnOneCell is True and after the learn cell was chosen. If you chose the previously learning cell, you should try to find the best matching segment otherwise you will always create a new one.

The code snippet is:

        if self.learnOnOneCell == True:
          # in learn on one cell mode, always learn on one cell per column
          # unless reset has just been called
          i = self.getSeqLearnCell(c)

        if not i:
          i,s = self.getBestMatchingCell(c,self.activeState['t-1'],
                                         self.distalDendriticInput['t-1'])

          if s is not None and s.isSequenceSegment():
            s.totalActivations += 1      # activationFrequency
            s.lastActiveIteration = self.iterationIdx
          else:
            # if best matching cell does not exist, then get least used cell
            i = self.getLeastUsedCell(c)


        else:
          # If we get back i, then we should find its best matching segment
          # if it exists
          s = self.getBestMatchingSegment(c, i, self.activeState['t-1'],
                                         self.distalDendriticInput['t-1'])

from htmresearch.

@subutai If this bug exists, the temporal memory won't learn and predict anything, right? I tested this, by introducing the bug into learn_on_one_cell_temporal_memory.py, and all the existing tests failed, because nothing was ever predicted. Do we need to explicitly test for this case, or are the existing tests enough?

(This is assuming that connectedPermanence > initialPermanence in the tests.)

from htmresearch.

Hmm, I see. All the old scripts used to pass without this change. This fix is needed only when you do slow learning. If our tests all use slow learning now, I guess it is fine to not make any additional tests.

from htmresearch.

Then maybe we can think of it as, "a good way to test this fix is to have a test for slow learning"? After all, that is the functional effect of this fix. As long as we have that, we're good.

from htmresearch.