Modeling of neocortical plasticity

I came across a nice summary of a few neocortical plasticity mechanisms that some deep learning researchers are trying to incorporate in their models.

Most of it reminded me of the HTM neuron described in the beginning of the BAMI document.
Some additional info may interest the community.

On a side note, the paper put an emphasis on Martinotti interneurons with a functional role that seems similar to the HTM spatial pooler. But I was thinking that the HTM SP was modeling the role of basket and/or chandelier cells, not Martinotti cells. Is it something different?

Martinotti neurons, which are activated by such cliques of pyramidal neurons, and subsequently inhibit pyramidal dendrites (Silberberg and Markram, 2007) provide well-timed inhibition to block further NMDA spikes (Doron et al., 2017), and put a limit on the maximal pyramidal clique size, but also suppress activation of competing cliques (e.g. Winner-take-all (WTA) dynamics).

The paper is available here (short read)


Looks like the author is searching for ways to model active dendrites in Deep Networks. That would be interesting for us, as well! As for the Martinotti cells, I’m not sure I understand the author.

I think they’re talking about competition for high frequency firing or something like that.

I’m not sure whether Martinotti cells come in multiple forms but I believe they only target distal apical dendrites. They are sort of spatial pooler-like because they target fairly broadly and are inhibitory. The distal apical dendrite of pyramidal cells can produce high frequency firing, and Martinotti cells themselves have facilitating inputs which means they respond more to high frequency inputs.


The extract gives some key differences between those 3 inhibitory cells.

Martinotti cells that mostly target the distal apical tree might selectively control NMDA spikes in distal apical dendrites; basket inhibition might selectively control the Ca2+ generated at the main apical branch of layer 5 (L5) and L2/3 pyramidal cells; and chandelier inhibition (targeting the initial segment of the axon of pyramidal neurons) could selectively control the somatic/axonal Na+ spike.

Martinotti cells target distal apical dendrites in L1.

Basket cells only target nearby pyramidal cells on their soma / proximal dendrites. I guess that this is the kind of inhibition that is currently modeled by the classic one-layer HTM Spatial Pooler.

When sufficiently activated, inhibitory cells prevent spike propagation to the axon terminals of the pyramidal neuron being inhibited. But contrary to basket and Martinotti cells, chandelier cells only target the Axon Initial Segment (AIS), thus allowing the back propagation of action potentials (bAPs). This difference probably have a role in learning.

1 Like

True. Inhibiting axon itself right after soma seems to lead to a scheme where neuron is considered “fired” for learning purposes, even if it doesn’t transmit any info further down.

However I don’t see the quoted explanation matching your model assigning basket to soma/proximal?
It talks about apical trunk, which imho has an effect on how the dendritic spikes originating in segments of the apical tuft would be integrated… iirc I read elsewhere that their effect could range from ignored, all the way to possibly driving, through weakly or strongly modulating.

also, that’s a lot of might and could

I didn’t understand that part. Are they connected to distal parts for their inhibitory output? In this quoted part you seem to imply that they take distal spikes as their input. Is this a two-way connection?

1 Like

True. It was based on other readings.
The focus of this paper was NMDA spike, so the soma/proximal dendrites were not part of the description I guess

1 Like

Yup. and assuming HTM 1000’s synapses paper as true, and linking that to your intuition about learning, I’d go for chandeliers as the TM inhibitors, since this would allow the SP to continue to wire similar proximal input to all cells in the minicolumn, as if all were “fired”, even if subsequently inhibited by a predictor-in-column.


Yes, it’s a two-way connection, for L5 thick tufted cells at least. Distal apical dendrites produce bursts of spikes. Martinotti cells have facilitating inputs, which might make them respond to bursts (or maybe not). If they do that, they might mediate competition for bursting. Or oscillations or ending bursting or who knows.

1 Like

I like your idea for the chandelier cells!

There are also other speculations on their function. Some of them were recently exposed in this video from the Allen Institute:

They recorded chandelier cells in vivo and showed that their spontaneous activity was synchronous between the cells, pointing towards a modulatory function. And because there is a gradient of the number of their axo-axonal synapses across the cortical depth, it creates a gradient of excitability of pyramidal cells in L2/3 during arousal states (less exitability in L2, more in L3). Interesting idea to have a modulable gradient of excitability in L2/3 !

A slide from the talk:

1 Like