This paper seems to consistent with the HTM model of dendretic spikes: https://www.cell.com/neuron/pdf/S0896-6273(19)30439-8.pdf
Since I can’t find a none paywalled version, this is a summary: http://news.mit.edu/2019/neurons-dendrite-role-computation-0606
7 Likes
“It’s an enormous, far-flung antenna that’s listening to thousands of synaptic inputs distributed in space along that branching structure from all the other neurons in the network,” Harnett says.
Antennae sure work for me! I was expecting neurons would (like on other types of cells for sensing things like motion and odors/chemicals) have them too, but I did not exactly know where they were, as though neural antennae were too small to image or something. Now I do! Thanks!
Although it’s just what I would (in a perfect world) for vocabulary streamlining and intuitive purposes would personally like to see happen in response to being confirmed true: after a phase in period where “dendrite/antenna” or “antenna/dendrite” is used to warn of two words for the same thing everyone switch to calling a dendrite an antenna. At least this way be easier for newcomers to remember which is sensing, the axon or dendrite/antenna.
2 Likes
MIT article says there are dendritic spikes at the “same time” as axonal spikes? I am guessing it’s for the lack of temporal resolution, it should actually be before: priming the soma, and after: synapse-reinforcing feedback? Anyone knows if dendritic spiking is bi-directional?
1 Like
I searched using “dendritic spiking is bi-directional” and although there was no exact match this one right away caught my attention. It’s what a patch of whiskers (and associated barrel cortex area) or antennae sensilla are expected to be good at sensing, in the signal pattern:
1 Like
Thanks! From what I gathered, “directional” here refers to motion of a stimuli, not of spiking. The paper refers to to the direction in a dendrite as pre-synaptic (feedback?) and post-synaptic (forward?). “Presynaptic DS is apparently necessary because it is more robust across the dendritic tree”.
“Postsynaptic inhibition generates robust DS in the isolated dendritic tips but weak DS near the soma”.
So, it’s not clear if the postsynaptic feedback is only diffuse inhibition, or there also are synapse-specific excitatory spikes?
There was this found in their model that indicates “spikes initiated within local dendritic regions” traveling both ways through their respective dendritic tree:
Our simulations show that sub-threshold PSPs from the distal dendritic regions of the On-Off DSGC are heavily attenuated by propagation to the soma, but that spikes initiated within local dendritic regions can propagate with high probability to the soma and back-propagate to the remainder of the dendritic tree. Therefore active amplification of DS appears to take place during spike initiation in the dendrites.
This seems close to what you are describing. If inhibition is optionally included then back-propagation can be stopped.
1 Like
Great, thanks Gary. But it’s weird that back-propagation is to remainder of the tree. It implies that adjacent but not spike-causing synapses are potentiated. Fire together, wire together would be reinforcing the spike-causing ones.
2 Likes
A weird idea keeps coming to me - that is that the fire-together/wire-together thing is partly or wholly a local thing based at the synapse level. I envision a mechanism where the products of the metabolism of firing interact with the extra-cellular chemicals to promote growth.
I extend this with the thought that some residue of this experience (a metabolic residue) makes the synapse susceptible to the soma based firing wave as the local back-prop training wave as envisioned by Numenta.
3 Likes
This new RNA related information just arrived by email from Camp:
We are now at a level of detail where things like that need to be modeled in.
So don’t stop there! Keep going on that thought. Almost everyone reading this is probably wondering “back-prop training wave what”? Best I can do is recommend asking Bitking for more details on what he knows about how that part works.
The predictive model assumes that some part of the dendrite arbor fires on a sensed pattern, making a prediction that we will fire again soon. The prediction is a partial depolarization but not a firing of the main soma body. This prediction is the central theme of the Temporal in HTM.
Later (but not much later) the main soma body is fired - a very big event on the cellular level, and the action potential fans out through through the output axon AND back up the dendrite tree. (The back propagation wave) HTM theory has an assumption that the synapses that predicted this firing are strengthened; this is a key element of the theory.
4 Likes
Thanks Bitking, that made easy sense of everything. The paper mentioned back-propagation but exactly how this related to HTM theory was best explained by you.
I’m now wondering why there is no online HTM neuron in a dish, lab, yet. Brush it a certain way with activity and soma fires. Does this sound like a relatively easy one to you too?
1 Like
Sorry, I can’t keep a cactus alive.
1 Like
After adding a supply of virtual microglial cells to nibble the buds just right we’re likely into 10,000 hours of work to get almost there, but we have to start somewhere.
Thousand Brains type thinking simplifies the problem by being possible to use metabolic network level SDR’s and whatever else is needed from HTM theory to approximate overall synaptic and RNA behavior that leaves metabolic residue or whatever. All by itself a single neuron is expected to have a small but impressive brain of its own that makes good predictions, but where it’s for that reason expected to not be easy to model it would be a shame if it were easy as an electronic logic gate. At least requires no watering at all.
TBT depends on this older HTM working as described; this is assumed as being in place.
As far as glial cells not covering synapses - that just makes sense. If they are to be exposed to axons to make contact it would not work very well if the were covered up.
With all theories being to some extent “tentative” it should at this early point in time be no problem to where need be take “HTM” right out of TBT. You may have already seen my somewhat related reply to Matt for the last hangout.
There only needs to somewhere in the cortical mass be found the whole bunch of smaller brains with a moving straw views of the outside world. In fact I was quietly hoping that the theory would end up needing to include at least some of the individual neurons. In that case TBT only becomes even more awesome.
I wouldn’t say this at all. The sequence memory algorithm we explained in our 2011 white paper is essential to all further Numenta theory. If we find that this model is inaccurate, we will have to rethink all the sensorimotor theory and TBT.
2 Likes
Yes, you can “rethink” your theory. That was more or less what I was saying. The whole thing does not necessarily have to be thrown out just because one part needs updating.
From my perspective TBT would be a shame to abandon now. For that reason Bitking’s mention of “TBT depends on this older HTM working as described” was a little bit alarming to me.
1 Like
It is one thing to adjust details of HTM sequence memory as we better understand biology.
It is another thing to throw the whole mechanism out because it is not working at all like we thought.
I have not seen any evidence of this 2nd thing yet, has anyone else? Even this paper that started this thread seems to confirm it.
1 Like
@rhyolight I don’t think that it is the issue - it’s just that the connection to the older papers get more tenuous as time goes forward and there is a large conceptual distance between today’s TBT and the papers it is based on.
I am always grounded in these older papers and the connections though to todays work but it does take a certain depth of understanding of the theory to see these connections all the way from the synapse mechanism level through the mini-column level all the way up to the impact on the higher level representation level.
This does go further, to the map level representation, onto to the map-to-map level, and global interactions between the maps and the sub-cortical structures. I have been working at these levels for many years so for me; theory tweaks at different levels always have to be considered through the lens of what impact it has on all these levels.
While this can be complicated it also has the advantage that known behaviors at various levels puts some constraints on possible answers at other levels. This sanity check does give a nice go-nogo check on proposals when I encounter them.
This does make it very hard to find someone to talk at a peer-to-peer level at parties.
3 Likes