Updating the spatial pooler when no cells win

[calumroy]

The potential synapses that connect a column to the input are only updated when a column wins.
This presents a problem when no column wins as no potential synapses permanences will be updated and no column will learn this pattern.

Sometimes no columns win when all the columns have learnt inputs already and no columns potential connected synapses connect to this new input. How should a column be chosen to update it's synapses so this new input can be learnt?

[calumroy]

A possible solution to the above problem.

The following requires changes to the overlap and spatial pooling learning algorithms.

• If a columns potential overlap score (the overlap calculated using the potential synapses not connected synapses) is larger then the minimum overlap value,
• the column is not inhibited by any other columns,
• the column was last activated the longest time ago with respect to the other columns (within the inhibition radius) that also adhere to the above points.
• The column has a higher tie breaker value then the other columns (within the inhibition radius) that also adhere to the above points.

Then activate this column as the winning column. Update the potential synapses that are connected to active inputs (similar to the current spatial pooling learning process). I'm unsure wether to also decrement synapses that aren't connected to active inputs (this needs extensive testing)?

[calumroy]

To implement the above solution the overlap score from all potential synapses for a column (apical dendrites) are required. Luckily the current theano overlap calculator class already does this and outputs it.

We also need to change the inhibition calculator in the following ways;

1. Currently it inhibits any columns with an overlap value of zero. We need to change this so columns that didn't have an overlap value larger then zero aren't automatically considered inhibited. After the inhibition step there may sometimes still be columns with a zero overlap that weren't inhibited by the active columns, we also need a way to tell which columns these are.
2. This list needs to be accessible from the inhibition class.
3. We can then check these columns potential overlap score instead and decide according to the above steps for this possible solution if these columns should be activated or not. This is possibly another task for the inhibition calculator.

[calumroy]

After adding the new function it is apparent that some problems still exist.

1. Temporal pooling occurs more easily for layers even when no temporal pooling should be happening for example with random input patterns. The reason for this is that columns that burst are then favoured to stay on for the next input as long as the input has a good overlap with the columns potential synapses.
   This may not be a problem as higher layers should inherently be more stable anyhow. Their is still a large difference in the amount of temporal pooling for known vs unknown patterns.
2. Columns easily over temporally pool a pattern. Eg a column may learn the pattern ABC but if that pattern ever burst during the B input for example a column will be favoured to also temporally pool over all the other B contexts because it was last active. E.g AB'C and AB''C (B, B' and B'' represent the 3 sets of cells in the column representing input B. They represent different contexts of input B).

A possibility is adding back in boost but being more selective of when boosting can allow a column to get the upper hand to win inhibition. This needs more thought and testing but i'm tired now.