I understand the idea of using the apical dendrite as an input for a feedback signal. How is the sign accounted for? Sometimes more is requested and sometimes less is requested. How is this done with a simple spike train as input?
Evidence suggests that this is handled by the lower structures presenting both positive and anti-positive versions.
I wondered the same thing and @gmirey offered an example from the visual stream. Perhaps he can offer this link again?
Hey there. Not sure what you’re referring to, @Bitking. If that of opposite pairs pathways for sensory input, some early ref. data about this can be found in the “Eye, Brain, and Vision” classic by Hubel&Wiesel
some chosen quotes:
But I’m not sure I understand precisely Ed’s question related to feedback here.
Care to develop, @Ed_Pell ?
gmirey, I am thinking about Blake Richards presentation at ICLR 2018 on using apical dendrites as a communications channel to send “nudges” down to hierarchy. But the nudge can be a positive nudge, i.e. do more of that, or a negative nudge, i.e. do less of that. But with one apical stem how does it know positive or negative and how does that cause the synapses of the basal dendrites to change stronger or weaker?
If there is some kind of “clocked” timing that can define early signal from apical versus late signal from apical then we have a sign. But I do not see where this timing info is coming from.
Ah. I get it. Well… then I’ve no paper to answer to that.
I’d simply point out that in that video he used some insight from biology to refine some NN model, which lived happily with both backprop and even possibly negative weights to this point… so his reliance on a sign maybe here, is not a strong clue that it indeed has a biological counterpart.
On the other side, why wouldn’t brain waves from thalamus be providing you those clocks you ask for, as sufficient info to either increase (pre then post) or decrease (post then pre) those synaptic connections accordingly, if that is what you aim for ?
Also. If your endeavor is to go from HTM to the handling of apical feedback… then increasing or decreasing those synapses on apical tuft, HTM-like, is maybe not too different in implementation from increasing and decreasing proximal and basal/distal themselves ?
But then, coming with an HTM model caring about apical and hierarchy is some kind of a grail at this point.
I’m not sure, but maybe he was talking about the inhibitory interneurons that are in L1, so feedback inputs to L1 can be excitatory or inhibitory. There’s also a circuit (activated by feedback input if I recall correctly) where one type of interneuron (VIP cells) inhibits martinotti cells which inhibit pyramidal cell apical dendrites, so maybe he was talking about that. The word nudges seems like a broad function, rather than controlling single cells.
Inputs from multiple sources tend to have a supralinear effect. Even if they just add, that can increase the chances* of reaching firing threshold a lot. So basal and apical dendrites might sort of multiply.
*not to imply this is because of randomness. It’s just easier to reach threshold. I
f by change you mean long term plasticity, I don’t know, but it might be related to how the cell fires anyway. Maybe he was talking about how apical input can cause cells to fire a burst of a few spikes at extremely high frequency, and an article found that bursts send a much better signal to distal basal dendrites. That article was wrong, according to another article which gave reasons.
I don’t know if this is what you are talking about, but there is a later feedback signal, whereas when the cell fires, it sends an excitatory signal to its apical dendrite. These are sometimes called early and late. There’s also a less late response caused by cells firing in response to sensory input and then activating synapses on other cells.
This is what I understood as well. Almost all the cortical layers are modulated (nudged) by L1. For example, a particular projection to L1 may cause selective inhibition in L5 through the apical dendrites. If you give excitatory input to L1 this means that less is requested from L5 because it is inhibited more. If you inhibit L1 then you request more from L5 because you will get more excitation from L5. Pyramidal cells of cortical layers can be effected by both inhibitory and excitatory input. So the sign here is the input type; positive sign (excitatory) leads to depolarization (easier to fire) and negative sign (inhibitory) leads to hyperpolarization (harder to fire).
By the way, @Casey is it possible for L1 to excite post synaptic targets among cortical layers? Do you know any studies supporting this?
Yes, although maybe just by their apical dendrites. Besides direct excitatory synapses, input to L1 can activate VIP cells (which might be in other layers) which inhibit martinotti cells, thereby disinhibiting apical dendrites. Martinotti cells and VIP cells both also inhibit PV cells, so things get more complicated at that point regarding proximal and basal excitation.
Sources for the circuit:
A disinhibitory circuit mediates motor integration in the somatosensory cortex (Soohyun Lee, Illya Kruglikov, Z Josh Huang, Gord Fishell, and Bernardo Rudy, 2013)
Whisker M1 -> whisker S1 L1 excites VIP which inhibits SOM cells likely disinhibiting L2/3 cells. in terms of bursting. The same circuit may exist for L5 TT.
Inhibitory Circuits in Cortical Layer 5 (Naka and Adesnik, 2016)
VIP cells inhibit PV and SOM (somatostatin) cells. VIP L2/3 cells also inhibit those in L5.
Sources which mention VIP cells
Distinct Roles of Parvalbumin- and Somatostatin-Expressing Interneurons in Working Memory (Kim et al., 2016)
Region-specific spike frequency acceleration in Layer 5 pyramidal neurons mediated by Kv1 subunits (Miller et al., 2008)
Somatostatin-expressing neurons in cortical networks (Joanna Urban-Ciecko & Alison L. Barth)
I cannot find a free version, unfortunately.
Feedforward motor information enhances somatosensory responses and sharpens angular tuning of rat S1 barrel cortex neurons (Khateb, Jackie Schiller, and Yitzhak Schiller, 2017)
A Computational Analysis of the Function of Three Inhibitory Cell Types in Contextual Visual Processing (Jung H. Lee, Christof Koch and Stefan Mihalas, 2017)
Signaling of Layer 1 and Whisker-Evoked Ca2+ and Na+ Action Potentials in Distal and Terminal Dendrites of Rat Neocortical Pyramidal Neurons In Vitro and In Vivo (Matthew E. Larkum and J. Julius Zhu, 2002)
Long-range recruitment of Martinotti cells causes surround suppression and promotes saliency in an attractor network model (Pradeep Krishnamurthy, Gilad Silberberg, and Anders Lansner, 2015)
Sources which mention somatostatin+ (slightly larger group than Martinotti) cells but not VIP cells
Parvalbumin and Somatostatin Interneurons Control Different Space-Coding Networks in the Medial Entorhinal Cortex (Chenglin Miao, Qichen Cao, May-Britt Moser, and Edvard I. Moser, 2017)
Global dendritic calcium spikes in mouse layer 5 low threshold spiking interneurones: implications for control of pyramidal cell bursting (Jesse H. Goldberg Clay O. Lacefield Rafael Yuste, 2004)
Layer 6 Corticothalamic Neurons Activate a Cortical Output Layer, Layer 5a (Juhyun Kim, Chanel J. Matney, Aaron Blankenship, Shaul Hestrin and Solange P. Brown, 2014)
Chrna2-Martinotti Cells Synchronize Layer 5 Type A Pyramidal Cells via Rebound Excitation (Markus M. Hilscher , Richardson N. Leão, Steven J. Edwards, Katarina E. Leão, and Klas Kullander, 2017)
A cellular mechanism for cortical associations: an organizing principle for the cerebral cortex (Larkum, 2012)
Enhanced dendritic activity in awake rats (Masanori Murayama and Matthew E. Larkum, 2009)
I cannot find a free version.
A Top-Down Cortical Circuit for Accurate Sensory Perception (Manita et al., 2015)
Anatomy and physiology of the thick-tufted layer 5 pyramidal neuron (Ramaswamy and Markram, 2015)
The organization of two new cortical interneuronal circuits (Jiang et al., 2013)
Feedback inputs also tend to go to layers besides L1, I think. Layer 5 slender tufted cells, at least in barrel cortex, receive proximal input from second order thalamus (POm, although more so from its subnucleus which receives both L5 and sensory input on the same cells and less so from the other subnucleus which only receives L5). Based on the response latencies, I think POm drives these cells.
Source for feedback to layers besides L1:
A Top-Down Cortical Circuit for Accurate Sensory Perception (Manita et al., 2015)
M2 (primarily L2/3/5a/6) -> S1, with boutons in all layers. I’m not sure this counts as feedback though, since they aren’t both in the what or where pathway.