Models of Neocortex behind HTM

Continuing the discussion from Need references on cortical mini-columns:

@Jonathan_Mackenzie I’m not clear what you mean by “the model of the mammalian neocortex used for HTM” in your question. HTM uses a totally non-controversial large scale description of the neocortex as follows:

The neocortex is divided in to Regions (eg V1, V2, S1 etc), which are spatially coherent patches of the cortical surface (grey matter), connected together by bundles of axons (white matter), and also connected to other parts of the brain.

HTM: we pretend these Regions are distinct, which is a reasonable simplification.

Each Region has a laminar structure, consisting of 5 or 6 apparent layers, L1-6, counted from the surface L1 downwards. L1 has very few cell bodies, and is mostly made of horizontal axon nets (which come mostly from other Regions) synapsing on the apical dendrites of deeper layer pyramidal cells (mostly L2/3 and L5), connected to their cell bodies by long vertical dendrite “trunks”. The other layers (L2-6) contain large numbers of neurons, with the majority of connections being within-layer and highly local. The majority of non-local connections follow stereotypical patterns as follows:

Sensorimotor or “lower-level” inputs predominantly go to L4 and L6/L5. L4 axons mainly go to L2/3. L2/3 axons mainly go to higher Regions, and also go to L5. L5 axons mainly go to higher regions and to motor centers in the subcortical brain. L6 axons mainly go to L4, L5, to the thalamus, and to L1 and L6 in lower Regions.

This video shows the laminar structure in mouse cortex. The green-labelled neurons are L5 pyramidal cells.

HTM simplifications: We pretend we have 4 distinct layers - L2/3, L4, L5, L6 (in reality there is no sharp transition between layers, and some people count many more sublayers). Each layer is modelled as a 2-d array of minicolumns of pyramidal neurons. A Layer in NuPIC corresponds to one of these idealised layers, typically L4 or L2/3.

There is no non-controversial understanding of exactly what all these layers do, and how they cooperate. A simplifying idea is to follow the grossest information pathway based on the predominant inter-layer connection maps. This pathway is Sensory->L4->L2/3 (-> higher region) ->L5->higher region, combined with a feedback pathway (higher region L6)->L1->L2/3 and L5, and an attention/modulation/control pathway (higher region L6)->L6->L4 and L5.

Numenta is currently working on the L4->L2/3 bit of this, which involves mainly Temporal Pooling by L2/3 cells of transitions and sequences in L4 SDRs. In addition (as far as I can gather), the L2/3->L5 and L1->(L2/3 and L5) motor integration components are being researched.

I proposed a theory which describes the multilayer Region from a top-down functional/mathematical point of view, treating each Layer as a Dynamical Systems Modeller, and describing how the Layers fit together in a local network. This idea uses plain HTM Layers as components. Discussion of and link to my paper is here.