Is there any information available on a systems understanding of the Neocortex, ideally using a metric based systems engineering framework? For example, how a particular neuron type acts as a system, its performance, resource and functional requirements, its dependencies to other systems and subsystems, the Cortical column as a larger system, etc
I have been pursuing this same thing for a long time.
I have been posting information on various levels at different times.
At the lower levels this paper is about as good as any I have seen:
Cortical Neurons and Circuits: A Tutorial Introduction - Richard B. Wells - April, 2005
Moving up to the mini-column level this paper is an excellent resource:
The columnar organization of the neocortex - V B Mountcastle - 01 April 1997
This post covers the larger dimension of spacing between the mini-columns:
I have spent some time on the forum trying to bring out the system level behavior of the inter-column connections and emergent signalling - perhaps best covered in this rather long post:
Moving up another level, map to map connections and the overall organization of the entire cortex, another very long post:
I have been nibbling on the lower brain structures for a long time but I don’t have a single post that ties them all together. The problem is that they are not a regular as the cortex to each structure really needs its own detailed section.
I would be delighted to see this all combined into a single rigorous resource.
I have collected pointers to much of my exploration of various part of the brain in this post and related thread. I don’t think it is as rigorous as you are seeking but at the time I did the individual posts I was not thinking of these things as parts of an overall system - I have been thinking about each in isolation. Some of these posts dive into various sub-cortical structures:
Fantastic - thank you for the references.
A thing to think about: resources required.
If you think of each map as the same as a graphics plane, with about 100 or so in the brain you can use this to visualize what a functional architecture might look like.
Assume for a starting point that each map is 1000 x 1000 nodes. Perhaps the same as a mini-column.
Each mini-column has perhaps 100 cell bodies, each with at least 10 each, proximal and apical dendrites. I will leave it to you to decide how many synapses each dendrite should have, but JH suggests that it could number in the thousands.Due to the important property of sparsity only a few percent of that number would be active at any time the number does dictate memory requirements.
If you shoot for brain-like speeds you will have to process all nodes at least 10 times a second. I have reason to think that the operations to resolve sparsity competition within a map should run at 40 Hz.
The output is about the speed as a ordinary video frame, and a good GB speed router should be able connect the right map processors to move the “finished products” of the local processing. There is no need for any map to have direct access to another maps local processing memory.
Graphics processors can pump out frames at 100 Hz rates so some of the questions revolve around how well you can mate these processes to the kinds of things a GPU does OR how good you might be at designing your own ASIC.
I am assuming that the processing maps to one boxen per map, each with some sort of GPU like processor. If it were possible to make an off-lease workstation run each map, at $300 each, you could build a brain for $300 x 100 or about 30 grand. Not pocket change but also - not a moon-shot kind of thing. It would not require Google or Microsoft (or DOD) resources to make it happen.
As the GPU makers start to shift the designs to be better at neural network tasks this scenario gets ever more doable.
This scheme still leaves the sub-cortical structures as an open question.
I think unlike typical AI approaches that brute force compute through massive amounts of data, a hierarchical approach, especially using SDR, can be extremely efficient. So that a single cortical column 5 levels up the hierarchy could be encapsulating a massive amount of learning, but dealt with in a “simple” way. There is also an idea that neurons (or dendrites) can do multi-processing: https://actu.epfl.ch/news/the-way-a-single-neuron-processes-information-is-2/ .
I’ll be working on a systems relationship graph and posting here as it evolves in the next few weeks…