Considering the limbic system or cortex in isolation really misses the point. As I said earlier - you have to look at it from a system point of view to make any sense of this whole thing.
Some points to consider before I try to put it all together:
- What exactly does it mean for a section of the cortex to “recognize” something?
I know that the Numenta people are putting a lot of effort to get the backward flow to be some kind of location signal so that the sense of touch can be locally combined with a mysterious location signal so that all your bits of fingertips can say - CUP! So what listens to this CUP message? How does the “listener” know what to do with the CUP message?
- This might be a bit of a stretch but ask the same question about the fibers from the limbic system. Either we accept the concept that cortex is cortex or we have to say that different parts of the cortex work in different ways. Does the limbic system have changes that we can detect? Can you turn it over and feel edges? That sounds kind of silly to say it and I think that it is just silly when I try to fit it into any kind of workable model.
- Now try the same thing with the visual system. The middle of V1 is somewhat the same as the bits sensing the tips of your fingers. For the visual cortex, we have some very good research that documents that it is sensing edges and such but that does not seem to have models of CUP in every bit of the visual field.
Maybe it works differently?
Let’s Identify some of the system components.
First - the larger structures
- Sensory cortex.
- Association areas (hubs)
- All the bits of cortex between the sensory areas and the hubs.
- Limbic system - thalamus.
- Major fiber tracts. - sensory to hub direction. (both the WHAT and WHERE streams.)
- Major fiber tracts - frontal to sensory direction.
On a more local level
- Layer II Grid-forming cells.
- Inhibitory cells with a grid scale effective range.
- Layer IV temporal-sensing cells.
- Inhibitory cells with a column scale effective range.
For this introductory chat, I am skipping many other system level components that I feel play important parts such as the hippocampus/cortex interface (entorhinal cortex), hippocampus, RAC and the pulvinar. I will try to paint a fuller picture with these and other parts later. That will cover attention and learning including one-shot learning
Talking local but thinking global - the pyramidal cells genetic programming can control the branching density, reach, and targets of each of the major dendrite groups: proximal and distal. The cells can also control the targets of the axons.
By varying these parameters you have the same pyramidal cells but very different behaviors.
The connections of inhibitory cells further tailor this behavior. Taken together, these variables give at least the two types that I will use in the following discussion but I am not excluding the probability that there are other highly useful behaviors in various regions of the cortex.
temporal-sensing (Numenta model) cells have distinct populations of the proximal and distal clusters. The long neck of the distal cluster allows the cell to be biased to firing faster giving the predictive state. The axon projects to the limbic system. There are interconnections with the grid-forming cells and short-range inhibitory cells. These inhibitory cells keep the action local - within the column. This allows the various cells within the column to compete to say that their prediction of the future is the best and suppresses other cells in the column.
Grid-forming cells are the communication specialists. The shorter distance to the distal fibers means that the sensed pattern has no time delay properties - these cells are simple pattern sensors. The communication output is both grid range inter-cell communications and cortical map-to-map signaling. This grid range cell to cell activity is shaped by the longer range inhibitory cells. The competition here is for the grid forming activity. See my HTM Columns into Grids post for more details.There are also connections to the temporal sensing cells.
On a slightly longer scale but still within a single map we have the sensory cells being bombarded with stimulus but no coordinated activity yet. As the real world comes in all of the layer II pattern sensors are all competing to recognize some sort of pattern. As the BAMI paper points to the trained SDRs embedded in the dendrites are like a key that only matches the pattern that they have trained to match. The possibility of matching some other pattern is vanishing low. If the cells do match some bit of a pattern they give the neighbors at “grid range” a kick. If that neighbor is also matching on a pattern it kicks back. Note that this mutual reinforcement will be VERY strong if a large population of cells are all seeing parts of a pattern that they have all learned - rapidly establishing a grid with the related inhibitory cells smothering competing patterns. This also produces a naturally sparse pattern.
Going Map to map - the early stages The temporal cells are excited by their companion grid cells and start predicting. This fires down to the limbic system. Things start to get more interesting now. The thalamocortical reciprocal connection comes back to layer IV and sets up the thalamocortical resonance cycles. Lateral connections in the thalamus spread this activation over the entire area of the grid activation area forming a pool of recognition. This excitation is now propagated to other maps via two pathways:
One is a general activation signal through the thalamus, the second is the projections from the grid-forming cells. The thalamus connection is a simple tonic priming the receiving map with a specific temporal wave to be in sync with the very specific message from the grid-forming cells.
Map to map spreads out This activation of synchronized activity ripples from the original sensory area - relatively weakly at this stage - on it’s way to the sensory hub in the parietal lobe. This projection of axons from the grid-forming cells is likely to be spatially coherent but it is possibly spread out to “smear” the pattern over a larger space. If you are familiar with the FFT process time is converted to space by the arrangement of sampling that groups related signal together by some temporal relationship. It is possible that the same sort of transformation is performed by the arrangement of interconnecting fiber bundles. For example - the spacing of the projecting fibers could give different spatial scaling to different target maps.
Meanwhile - in the limbic system the various need sensors in the body are sorting out which need is the most important. The dozens of interconnected clusters are competing to signal that what they want is the most important thing that needs attention. These cells are more like bundles of oscillators. The needs are not a static pattern but a dance of activity that the frontal cortex registers as an activation pattern. This is the sensory stream that the frontal lobe is sensing and has trained on. This activity ripples to the hub of the frontal lobe.
Meeting of the minds! These hubs (raw sensory, needs, parsed sensory) have long reciprocal connections. These connections of activity go up and down the paths adding very strong activation energy to patterns that match; at a minimum, there is twice as much activity for matching patterns! If the needs of the frontal lobe match up with the sensed environment there is a strong reinforcement that triggers a Global workspace Ignition[1]. Various “side chains” of connections shape this workspace but the counter-flowing river of reinforcement raises this pattern to the level of awareness and tend to suppress all other activity.
Perhaps you are familiar with the cocktail effect. This is a very good way to understand what this reinforcement is like - it is the most accessible part of this process. As you listen to a conversation in a noisy room the recognition of a single stream of words sharpens and shapes your perception stream. As your brain tunes to this pattern, all others are filtered out.
Now what? In the frontal lobe - the best match for a pattern is reinforced and elaborated to some activity. This may be internal to the brain (thinking) or some motor program (acting). Some of these motor programs can be directing the eyes (visual attention) or extracting some sound (listening) As I have stated before - one of the side chains of activity is the loop from the early motor stages to the corresponding sensory stages that form your internal consciousness loop. This Global Workspace can chain to other related activity patterns in a continuous process as the needs sensors continue to prime the frontal cortex. (goal directed activity)
This is so much more but I will stop at this introductory system description.
[1] The Global Neuronal Workspace Model