Temporal pooler isn't robust.

[calumroy]

Temporal pooler brittleness.

Currently (current commit 7a07099) the temporal pooler works by keeping the first set of active columns on longer throughout a sequence to pool over the sequence and increase temporal stability. Once a sequence has been temporally pooled this method does not include any columns active other then the first ones of that sequence.

An example;
Sequence A,B,C,D is learnt and temporally pooled over. The columns making up pattern A would stay active through out the sequence. This works fine except when noise is introduced or the temporally pooled pattern is to be compared to other similar patterns.

Similar temporally pooled patterns produce different output SDR's.

The output of the temporally pooled pattern for A,B,C,D will be 100% different to the output of the temporally pooled pattern B,C,D. This is bad since both patterns are very similar and should create a similar output SDR.

Introducing noise disrupts the temporally pooled patterns significantly.

If a single input of the above pattern A,B,C,D contains noise then the output temporally pooled pattern can vary wildly. E.g If the unlearnt pattern A,E,C,D would produce a completely different output SDR after the input E since this was unexpected causing new columns to burst. What should happen is the unexpected input E should cause bursting but then the next input C should not cause bursting as the temporally pooled pattern A,B,C,D should have been expecting this.

[calumroy]

Numenta's third version of the temporal pooler

Numenta has produced a third version of its temporal pooler which aims at fixing these problems. See the github wiki at https://github.com/numenta/nupic.research/wiki/Overview-of-the-Temporal-Pooler

Basically numenta solves the problem by having columns staying active longer and longer through a sequence as well as having later columns become active earlier and earlier. This eventually means that for the above example the columns representing A,B,C,D start to activate closer and closer together. Once temporal pooling is complete then the columns for A, B,C,D all become active at the start of the sequence. The inhibition stage would then only allow a certain fraction of the columns from each pattern to be active.

Numenta's temporal pooler does a few things to achieve this robustness;

1. The columns have a union pooler which updates a "persistence value"; basically if a column is is active then check it's synapses and depending on the amount of proximal synapses that are active and predicted increase this value (they use a sigmoid function for this). The value is used in the learning process to allow columns that have a high enough persistence value to learn the next input (update the potential proximal synapses connected to active predicted inputs). NOTE: This does not activate the columns but increases the proximal potential synapses connected to the next input, so the next time that this pattern occurs the column can stay active throughout the input sequence for longer.
2. A columns potential synapses (potential proximal synapses) that are connected to inputs that were predicted and active in the previous time step have their permanence values incremented. This is called backwards learning. It is what makes columns become active earlier and earlier in a sequence.
3. Cells within a column can form distal synapse connections to a columns union pooler. Basically if a column has a large enough persistence value then a cell can form a connection to that columns UP. To the cell the columns UP just looks the same as any other previously active cell that normally a distal synapse would form a connection to. This has the effect of allowing cells to be in the predictive state when they have good connections to a columns UP which has a high persistence value. In the situation outlined above the input A,E,C,D would not cause cells to burst on the input C because those cells would have been in the predictive state due to column A having a high persistence value and the cells in column C having distal synapse connections to this keeping those cell in a predictive state even though there was an unexpected input E.