How new segment is created in winner cell in burst column for the first time stamp

bipul_kalita · October 5, 2018, 3:46pm

As in pseudo code of function burstColumn(column):
…
…
33. winnerCell = leastUsedCell(column)
34. if LEARNING_ENABLED:
35. learningSegment = growNewSegment(winnerCell)
…
…
lets say for the first SDR as input to Temporal Memory algorithm, first it selects an active column from SDR as active(burst) and randomly selects a winner cell , now I am confused about how segments will be created in that winner cell? since for now it’s only one winner cell in one selected active column for the first time stamp.
honestly I want to know how growNewSegment(winnerCell) function works.
thank you.

rhyolight · October 5, 2018, 5:17pm

Following the source code:

github.com

numenta/nupic/blob/master/src/nupic/algorithms/temporal_memory.py#L528-L535


@classmethod
def _burstColumn(cls, connections, random, lastUsedIterationForSegment,
                 column, columnMatchingSegments, prevActiveCells,
                 prevWinnerCells, cellsForColumn,
                 numActivePotentialSynapsesForSegment, iteration,
                 maxNewSynapseCount, initialPermanence, permanenceIncrement,
                 permanenceDecrement, maxSegmentsPerCell,
                 maxSynapsesPerSegment, learn):

github.com

numenta/nupic/blob/master/src/nupic/algorithms/temporal_memory.py#L628-L630


segment = cls._createSegment(connections,
                             lastUsedIterationForSegment, winnerCell,
                             iteration, maxSegmentsPerCell)

github.com

numenta/nupic/blob/master/src/nupic/algorithms/temporal_memory.py#L664-L692


@classmethod
def _createSegment(cls, connections, lastUsedIterationForSegment, cell,
                   iteration, maxSegmentsPerCell):
  """
  Create a segment on the connections, enforcing the maxSegmentsPerCell
  parameter.
  """
  # Enforce maxSegmentsPerCell.
  while connections.numSegments(cell) >= maxSegmentsPerCell:
    leastRecentlyUsedSegment = min(
      connections.segmentsForCell(cell),
      key=lambda segment : lastUsedIterationForSegment[segment.flatIdx])


    connections.destroySegment(leastRecentlyUsedSegment)


  # Create the segment.
  segment = connections.createSegment(cell)


  # Do TM-specific bookkeeping for the segment.
  if segment.flatIdx == len(lastUsedIterationForSegment):

This file has been truncated. show original

sheiser1 · October 5, 2018, 10:18pm

A new segment is grown on the winner cell chosen in each column (1 on each winner of the 40 activated columns)

bipul_kalita · October 6, 2018, 5:27am

so for the first time a new segment will be grown in the selected winner cell in 1 column from 40 columns(active SDR columns). But how many synapses will be there? because to grow a segment means to grow its synapses.

Paul_Lamb · October 6, 2018, 4:34pm

Each new segment will start out with up to maxNewSynapseCount synapses.

github.com

numenta/nupic/blob/master/src/nupic/algorithms/temporal_memory.py#L626




    nGrowDesired = maxNewSynapseCount - numActive(bestMatchingSegment)


    if nGrowDesired > 0:
      cls._growSynapses(connections, random, bestMatchingSegment,
                        nGrowDesired, prevWinnerCells, initialPermanence,
                        maxSynapsesPerSegment)
else:
  winnerCell = cls._leastUsedCell(random, cellsForColumn, connections)
  if learn:
    nGrowExact = min(maxNewSynapseCount, len(prevWinnerCells))
    if nGrowExact > 0:
      segment = cls._createSegment(connections,
                                   lastUsedIterationForSegment, winnerCell,
                                   iteration, maxSegmentsPerCell)
      cls._growSynapses(connections, random, segment, nGrowExact,
                        prevWinnerCells, initialPermanence,
                        maxSynapsesPerSegment)


return cellsForColumn, winnerCell

sheiser1 · October 7, 2018, 11:43pm

If there is a best matching segment on any cells in the column, the number of new synapses grown on the segment = ‘maxNewSynapseCount’ (default=20 I think) minus the number of active synapses already on that segment.

If there is no best matching segment on any cell in the column, the least used cell in the column (with fewest existing segments) grows a new segment with ‘maxNewSynapseCount’ synapses formed to the ‘prevWinnerCells’.

Topic		Replies	Views
Temporal Memory: handling no previous winner cells? NuPIC sequence-memory	8	1095	August 18, 2016
HTM learning algorithm - temporal memory Getting Started question	17	993	August 25, 2021
When there are 2 or more predictive cells in a column that becomes active are both chosen as winner cells? Implementations	8	725	May 5, 2021
Correct determination of winner cell in minicolumn NuPIC Community Fork sequence-memory	2	518	January 8, 2020
Nooby questions for Implementation Getting Started	6	700	May 5, 2020

How new segment is created in winner cell in burst column for the first time stamp

Related Topics