Union Pooler Psuedocode

Union Pooler Algorithm Pseudocode

Note: This algorithm is a work in progress. (5.11.15)

Data Structures

activeInput ­- Binary vector representing active cells from Temporal Memory


predictedActiveInput - Binary vector representing correctly predicted active cells from Temporal Memory

poolingActivation - A vector of scalar activation values for each cell in the Union Pooler

Parameters

activeOverlapWeight - ­Weight given to overlap due to activeInput

predictedActiveOverlapWeight ­- Weight given to overlap due to predictedActiveInput

activationFunction -­ A function that updates a pooling activation value based on an excitation or a decay. Possible forms include linear, exponential, or logistic.

maxUnionActivity -­ Maximum allowed size of the union SDR as a percentage of the number of Union Pooler cells

Algorithm

# Compute current active cells based on current input
activeOverlaps = calcOverlap(activeInput) 
predictedActiveOverlaps = calcOverlap(predictedActiveInput) 
totalOverlap = (activeOverlaps * activeOverlapWeight + predictedActiveOverlaps * 
                predictedActiveOverlapWeight) 
boostedOverlaps = boost * totalOverlap if learn else totalOverlap 
activeCells = inhibitColumns(boostedOverlaps) 

# Perform Spatial Pooler learning algorithm with boosting
adaptSynapses(activeInput, activeCells)
updateDutyCycles(totalOverlap, activeCells)
bumpUpWeakColumns()
updateBoostFactors()
if isUpdateRound():
    updateInhibitionRadius()
    updateMinDutyCycles()

# Update Union SDR based on new set of active cells; decrement pooling activation of all cells
decayFunction.decay(poolingActivation) 

# Add to the poolingActivation of those active Union Pooler cells receiving active inputs
addToPoolingActivation(activeCells, activeOverlaps)

# Same for Union Pooler cells receiving active-predicted inputs
addToPoolingActivation(activeCells, predictedActiveOverlaps) 

# Compute the current most salient cells in terms of poolingActivation. 
# Cells with zero poolingActivation cannot win.
unionSDR = getMostActiveCells(poolingActivation, columnCount * maxUnionActivity)
return unionSDR