L5 Precise Timing and Reducing Cell Count Requirement

This is a suggestion based on L5 tall bursting (TB) cells. There are three types of pyramidal cells in L5 [1], but this suggestion is only based on TB cells. TB cells are interesting because they can project to thalamic matrix cells and central pattern generators [2]. In addition, Jeff Hawkins proposed that thalamic matrix cells, which project to L5 [3], are involved in precise timing [4].

TB cells have two firing modes, single-spike mode and burst mode, where the cell generates several spikes in extremely rapid succession. I would like to suggest that a burst indicates a precise timing.

Without a previous element in the sequence, there is no reference point for the timing. As such, bursting also represents that the cell belongs to the current sequence. This is quite advantageous because it allows L5 to use far less cells.

In temporal memory, if each cell in a column were sensitive to sequence context and timing context, each column would require hundreds of cells because number of cells per column = number of buckets for sequence context * number of buckets for timing context. Instead, each column can simply indicate that it belongs to the current sequence, which is sufficient data for temporal pooling by the next region and may aid temporal pooling of unordered sets. TB cells can leave sequence tracking to cell types without extremely precise timing.

-L5 TB cells use bursts to indicate that the previous element of the sequence led to the current one after an extremely precise duration of time.

-Bursts therefore indicate that the cell belongs to the current sequence, telling the temporal pooler that the cell represents a continuation of the sequence.

-Sensitivity to extremely precise timing would normally require many cells per column, but ignoring the location in the sequence solves this issue.

Keep in mind that I don’t really know how to read studies, so I may have misunderstood the sources and chosen a biased interpretation. It’s also easy to assume that something which doesn’t fit serves a separate unknown function or is a result of unknown latencies and such.

TB cells can enter burst mode as a result of the distal basal dendrite or the distal apical dendrite. My notes on the distal basal dendrite are incomplete and contradictory, so I will focus on the distal apical dendrite. The distal basal dendrite appears to propagate burst mode [5] because it only responds to extremely high frequency input [5], so it may not be as essential as the distal apical dendrite. However, one study found that the distal basal dendrite responds to low frequency input in juvenile animals [6].

The distal apical dendrite normally only causes a burst when the proximal basal dendrite generates a spike and it receives sufficient excitation from L1/2 [7]. It can generate a burst with heavy excitation [7], but that level of excitation may only occur in experiments.

After the proximal basal dendrite generates a spike, a backpropagating action potential (bAP) propagates to the distal apical dendrite [7]. If the bAP arrives within a few milliseconds of sufficient excitation of the distal apical dendrite, the cell generates a burst [8]. As there is a delay between a presynaptic spike and the excitatory postsynaptic potential due to distance, TB cells may require prediction rather than correlation to burst. Furthermore, the window for coincidence detection is wider when the input comes before the output, unlike the other layers [9], supporting the conclusion that the distal apical dendrite receives input from neurons representing a previous element of the sequence.

The distal apical dendrite is compartmentalized [10] like the dendritic segments in temporal memory. As such, each TB cell represents a set of timings. Just like in temporal memory, this works because the whole population of cells unambiguously indicates the timing.

[1] http://cercor.oxfordjournals.org/content/early/2014/11/18/cercor.bhu268.full.pdf?cited-by=yes&legid=cercor;bhu268v1
[2] https://www.researchgate.net/publication/280060329_Anatomy_and_Physiology_of_the_Thick-tufted_Layer_5_Pyramidal_Neuron
[3] http://www.ncbi.nlm.nih.gov/pmc/articles/PMC4889215/
[4] http://lists.numenta.org/pipermail/nupic-theory_lists.numenta.org/2016-March/003564.html
[5] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3850222/
[6] http://www.jneurosci.org/content/26/28/7424.full
[7] http://www.jneurosci.org/content/26/41/10420.long
[8] http://www.nature.com/nature/journal/v398/n6725/full/398338a0.html
[9] https://www.researchgate.net/publication/221716382_The_Time_Window_for_Generation_of_Dendritic_Spikes_by_Coincidence_of_Action_Potentials_and_EPSPs_is_Layer_Specific_in_Somatosensory_Cortex
[10] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4481152/