Functional clustering of dendritic activity during decision-making


#1

This paper has huge implications on what is going on in the dendrites.
The claim is that much smaller dendrite segments than previously recognized can individually process task specific memory.


#2

This is a similar finding to that of the article discussed in thread:

In that article they find that nearby synapses cooperate with each other, and that this cooperation decays exponentially with distance along the dendrite. They relate the dendrites segments to function compartments. They find that by default the dendrite functions with a few large compartments but that certain conditions (such as inhibition) can split the dendrite into many small compartments. Bitking’s article finds that in-vivo, dendrites typically operate with many very small compartments.


#3

This is one of the things I’ve been looking into a little bit. Without having made a scientific exploration of the subject, I can still say that it generates interesting results.

You pay in computation and memory because the order of synapse creation needs to be preserved and I haven’t found any really effective way to evaluate dendrite activation without trashing the cache.


#4

I found that the state of the dendrite can be computed quickly if you assume that the effect of each synapse on the rest of the dendrite tree exponentially decays with distance along the dendrite. All of the exponentials factor and you can compute the end result in linear time (with respect to the size of the dendrite). In theory O(N) is fast…

Scientific code is designed to break the cache, it often uses each memory location once or twice. RAM on the other hand is reaching a price point where its no longer a limitting factor :slight_smile:


#5

Thanks I came here to ask about the same thing since I saw the article and it made me nervous that the problem could be even much more complex. But if I understood your reply, you already explored modeling that feature with modified HTM though and found it doesn’t blow up the computation to account for it? Will the main HTM implementation be changed now?


#6

We have no plans of this at the moment. Research is heading towards an attempt to explain orientation as a part of object recognition.


#7

So i found that it performed as well as the original model and took a lot longer to run. Maybe if i added up-down states or SOM inhibition to the model then the dendrites performance might improve enough to justify the added complexity of modeling at the 1um resolution Vs. 100um. You can read my report on this project at https://github.com/ctrl-z-9000-times/dendrite_experiments/blob/master/dendrites.pdf


#8

Thanks for your work on it. I get the uneasy feeling what you mention about “up-down states or SOM inhibition” might be necessary then? If this feature distinguishes human neurons it seems likely it is functionally relevant to the correct model. But I feel optimistic knowing people like you are already exploring that direction.

I hope the model will improve and someday they will make custom chips that implement genuine functionaly cortical columns like we do with GPU cores now!!