is it clear that proximal input of a layer comes just from the lower layer, or it might be coming also from the same layer? just like the way modulatory input comes from both the same layer and neighboring layers?
logically having a recurrent proximal input will have benefits e.g. :
the sequences AAA and AAB produce no shared active cells, cause they have no shared active mini-columns in the last timestep (think of it as location), but having the proximal input not just from the lower layer, but also from the active cells of this same layer, would make a part of the proximal input (state of the layer in the previous timestep) similar, which leads to shared active mini-columns that would create shared active cells. this way there would be a similarity at the end of the sequence.
Yes/no neuroscience questions usually have ambiguous answers, I think. That can be annoying, but at least it’s not like looking something up in a dictionary. There’s room for realizing you were thinking about a problem wrong for a long time, and you have to correlate evidence and resolve contradictions. That can be fun.
Here’s some evidence for recurrent proximal connections. My guess is that 1/2 of experimental results are wrong (or at least don’t apply to awake intact brains), and 4/5 are misleading without considering other things going on.
“Synaptic efficacy and reliability of excitatory connections between the principal neurones of the input (layer 4) and output layer (layer 5) of the neocortex” (Dirk Feldmeyer and Bert Sakmann)
Inhibitory interneurons probably matter a lot for minicolumn states.
I’m not sure recurrent proximal input automatically leads to recurrent minicolumn states, since the recurrence would probably have to activate all cells in each minicolumn the same way. Although, recurrence between whole minicolumns might not be necessary, since the whole minicolumn isn’t on anyway. If whole-minicolumn recurrence is necessary, such as perhaps to represent CAA similarly to AAA, I think it would have to involve interneurons. Some types of interneurons have recurrence with excitatory neurons and/or other interneurons, I think.
To make the state similar based on the prior state or prior input state, the recurrence probably requires a delay or a recurrent signal which outlasts whatever sends that signal. As I understand it, the distal connections in temporal memory produce a sustained signal using NMDA distal dendritic responses, which last maybe 100-300 milliseconds, I think. There are other options, like metabotropic receptors, which can produce excitation or inhibition lasting seconds.
It looks like they recorded the soma only, but I’m not sure. Also, distal dendritic input can produce responses at the soma, and it can influence the response to proximal inputs. Another thing to keep in mind is that some layers contain multiple types of pyramidal cells, so proximal input might not be recurrent.
If i understood correctly this and this youtube lectures from
(the The Brain and Computation Boot Camp at simons institute) Bartlett Mel from USC says ~75% of connections to pyramidal neurons are from other pyramidal neurons - much less from sensory inputs.
That may be correct, but that’s not really what the OP is asking. The question is whether a layer gets recurrent input (it’s cell state is processed as proximal input. And we’re saying we really just don’t know, but there is no reason not to experiment with this, IMO (meaning you the community, not Numenta, we are not planning on this).