Where does HTM theory incorporate logic gating at the dendritic arbors

[Tony]

I’m new to the HTM theory and still trying to get my arms around it.

I’m currently trying to find out where in the HTM theory the use of dendrite segments feeding into dendrite arbors and being processed as though they are logic gates is discussed. I’ve read a number of the HTM papers and haven’t yet been able to find it.

From reading the simulation code, it appears as though each dendrite segment is being processed as though it is directly attached to the soma. There doesn’t appear to be any interaction between the various dendrite segments of a neuron.

Any pointers to the appropriate HTM paper or clarification would be appreciated.

[Bitking]

I think we have a forest-for-the-tress thing here.
The connections that collect inputs to trigger an action potential ARE the gating.
The sets of connections form the SDRs that is described copiously in the Numenta literature.

The idea that these are some sort of dendritic logic element are the stuff of conjecture at this point and the sticking point is how some sort of logic would come to be.

[Tony]

I thought logic gates, operating at dendritic arbors, was considered “settled” science at this point. At least with regards to OR and AND gates. There was also a recent publication that generated some buzz about the ability for the ability of a single neuron to perform an XOR operation applying the same concept.

Just to make sure we are talking about the same thing, page 8 from this publication, https://doi.org/10.1101/690792, contains a nice summary of what I’m referring to. The image below is extracted from there:

[Bitking]

You are showing a proposed mechanism. I don’t recall seeing anyone showing this happening in the wetware.

This is what I can see a cortical column doing at this time:

What is the Basic Cortical Column Computation?

I struggle to pin down exactly what it is that a macro-column is doing; there is a lot going on here. Pattern matching (by pattern recall) Novelty detection (by pattern recall mismatch) Pattern learning (by novelty triggering learning) Anomaly signaling (by sequence mismatch mechanism) Sequence learning (by anomaly signal triggering learning) Prediction (by sequence learning) “Memory” (by the collective action of these mechanisms?) So we add these ensemble macro behaviors: Classifier (by pa…

I should add that the T of HTM is temporal memory, exactly the problem they are trying to solve in the referenced paper.

[Tony]

The image I showed was just to ensure we were talking about the same thing. It wasn’t intended as a reference.

Here is a reference for the XOR gating being done in the dendrite.

Dendritic action potentials and computation in human layer 2/3 cortical neurons

I don’t have any references at hand for the OR gating studies but do recall recently reading one recently which talked specifically about an invivo study where two separate dendritic spikes were observed “obliterating each other” in the dendritic arbor thus causing that junction to act as a logical OR gate. I’ll look through my tagged papers later to see if I have a copy of that paper still around.

Is it fair to assume that HTM theory doesn’t currently incorporate any computational logic at the dendrite segment level? I’m not referring to the coincidence logic currently described in their paper. Rather, I’m referring specifically to gating logic that would take place at the junction of two dendrite segments.

[Bitking]

It does not have dendrite and/or logic now but there is nothing to preclude extending the code that way.

I have real trouble working out how the synapse learning rules would work with this additional posterior logic. As it is, the learning is simple and local. Have you seen anyone that shows how the learning and activation rules would have to work with this posterior logic added to the picture?

[dmac]

Hi Tony,

The HTM model simplifies the dendrites. It makes it faster to compute, and easier to analyze with math. The HTM captures many key properties of dendrites.

Here is another model of dendritic arbors: https://doi.org/10.1016/j.celrep.2019.01.074 . It is realistic (conductance based) and fast to compute.

[Tony]

@Bitking

It does not have dendrite and/or logic now but there is nothing to preclude extending the code that way.

I have real trouble working out how the synapse learning rules would work with this additional posterior logic. As it is, the learning is simple and local. Have you seen anyone that shows how the learning and activation rules would have to work with this posterior logic added to the picture?

Currently I’m just trying to get a better understanding of what is included in HTM and what was abstracted out.

The big things that stick out to me as not being included are: 1) dendrite arbor gating, 2) timing delays associated with dendrite spiking/membrane capacitance, and 3) potential temporal information derived from spike trains.

I realize that much of this may have been considered during the development of the theory and discarded as just a function of the necessary biological infrastructure and unimportant to the core of the concepts on how AGI actually functions.

[Tony]

@dmac

Hi Tony,

The HTM model simplifies the dendrites. It makes it faster to compute, and easier to analyze with math. The HTM captures many key properties of dendrites.

Here is another model of dendritic arbors: Redirecting . It is realistic (conductance based) and fast to compute.

This looks like a really good paper that I haven’t seen before. I just added it to my reading list. Thanks for pointing it out!

[Bitking]

@Tony

The big things that stick out to me as not being included are: 1) dendrite arbor gating, 2) timing delays associated with dendrite spiking/membrane capacitance, and 3) potential temporal information derived from spike trains.

HTM focuses hard on the pattern recognition (temporal and spatial) at a higher level than these mechanisms.
The spike timing learning is notable absent and I feel that HTM would be richer for including this well known mechanism. In my own models I have included this but it is to early to say if it actually helps. I have not done much to feed it real data and evaluate what it does with it. In reality I am still working on getting my code in a state where I trust it.

[dmac]

@Tony

Currently I’m just trying to get a better understanding of what is included in HTM and what was abstracted out.

Inhibitory inter-neurons were also abstracted out.

The HTM operates at a large resolution of time. Each compute cycle corresponds to a whole brain wave.

[Tony]

@Bitking

@Tony

The big things that stick out to me as not being included are: 1) dendrite arbor gating, 2) timing delays associated with dendrite spiking/membrane capacitance, and 3) potential temporal information derived from spike trains.

HTM focuses hard on the pattern recognition (temporal and spatial) at a higher level than these mechanisms.
The spike timing learning is notable absent and I feel that HTM would be richer for including this well known mechanism. In my own models I have included this but it is to early to say if it actually helps. I have not done much to feed it real data and evaluate what it does with it. In reality I am still working on getting my code in a state where I trust it.

Once I feel as though I have a really good understanding of the HTM theory, I intend to code it as well. I’m making notes on those areas of abstraction that they removed, but I feel still add value, so I can turn them on or off as I run tests to see if they do add any value.