User:Graeme E. Smith/Collections/Model Series/Datamining/How your Memory Might Work

How Your Memory Might Work

‌I have written a whole book on this subject, if anyone is interested, called How Memory Might Work It was originally meant to be published before this book, but due to financial limitations it has not yet been published. Obviously I can't put the whole book into this chapter, but I do want to do a bit of an overview, so that you can see my model of memory and what it tells us about Intuition.
‌The brain is made up of Billions of Neurons with literally trillions of connections between them. In essence it can be seen to be one massively parallel neural network. It is made up of some 150 different types of neurons, if shape is any judge. It is anatomically mapped according to areas of similar neurons, and subdivided into organs by connecting tissue, and white matter which is often extensions of individual cells that are surrounded by Myelin, and en-mass form interfaces between distant neural organs. Each neuron can do three things, It can Store Memory, Transport that Memory from one end of the neuron to the other, and do some minimal processing. One reason there are so many different types of neurons is that they have different roles. For instance a long thin neuron with only one dendrite, and only one axion, would probably be a transport neuron, while a shorter bushy dendrite or Mossy Dendrite neuron would probably be more of a processing neuron, and a Neuron with a pyramidal Soma or Body, might be a Memory Neuron. Organs that are long and slim, tend to be associated with transport, and organs that are short and fat, or planar in configuration tend to be either processing or memory organs respectively. There are exceptions and no organ is limited to a single type of neuron, often being interpenetrated with many different types of neurons.

Pyramidal Areas in the Brain

‌The Memory Neurons tend to be found in the Pyramidal Areas of the brain, the Cerebral Cortex, the Cerebellum, the Hippocampus area, etc. As far as we know up until now, there are 5 main types of memory in the brain, and three of those types of memory depend on the Cerebral Cortex as their main storage element. The cerebral cortex, is a planar arrangement of neural tissues that wrap around the upper side of the brain, and makes up about 1/3 of it's mass by weight. To call the Cerebral Cortex a memory, is to ignore the fact that neurons pretty well have to do processing at the same time as they do storage. However the Pyramidal Neurons which are the main memory neurons in the brain, do very little processing themselves. Most of the processing involved is done by other neurons that surround and support them. The Cerebral Cortex is 6 layers deep, and 4 of those layers are memory neurons of one type or another, while the other two do processing and support roles.
‌There are essentially three memory loops through the brain one that loops through the Cerebral Cortex the Thalamus, and the PreFrontal Cortex, One that loops through the Cerebral Cortex, and the hippocampal area, and one that loops through the Cerebral Cortex, down the brain stem, involves the cerebellum, and back to the Cerebral Cortex again. These have been separated into loops associated with memory, skills, and indexing for Declarative Memory. It is interesting to note that projections for all three loops, terminate in the Cerebral Cortex, and gather in a part of the cortex called the Striate Cortex. From this it can be predicted that the Cerebral Cortex, being both the largest part of the system, and also part of all three loops, is in fact the center of the memory systems in the brain.

The Cerebral Cortex

‌In 1970 David Marr, a scientist at Trinity College in Oxford, described the Cerebral Cortex as "A self-classifying Content Addressable Memory" (Capitalization is mine) His characterization only really dealt with the first four layers of the cerebral cortex, the 5th and 6th layers have taken longer to understand. It turns out that theoretically there is only two ways to access memory, it can be Addressed using a place-code, such as computer memory is, or it can be accessed by content.It is important to know that while you can address place-code stored information by content, (it requires massive processing), you can't as easily access content addressable memory by place-code, especially if it is made up of Neural Networks. The problem is that neural network implementations of Content addressable memory exhibit something called phenomenal characteristics. Their memory is spread over the network of neurons, and does not gather in any particular locale to be read from.

Columnar Architecture

‌It turns out that the 5th and 6th layers of the Cerebral Cortex, are in fact a work-around for this problem, needed because you need a place-code addressed form of memory to implement Demand Memory, the type of memory where you ask for a memory by topic, and it is retrieved for you. The 5th layer acts as a sort of addressing array, that gathers together hundreds of neurons into clusters called mini-columns that can be pre-activated by projections from the thalamus, and the sixth layer gathers the mini-columns into a larger column architecture which localizes the output so that the Cerebral Cortex learns to store its memories at the columnar level. What this seems to create is a form of Neural Grouping where outputs of neurons within the group are suppressed and one centroid neuron fires for the whole group.

Dual Mode Cortex, Implicit and Explicit Memory

‌What this means is that the Cerebral Cortex can be addressed either by content or by place-code allowing two modes of addressing for the same memories. When addressed by the content, the memory is implicit, meaning it only responds to the content of the memories, but when addressed by the place-code it responds not only to the contents of the neural group designated by the place-code, but to the pattern that creates in the implicit memory, and the output comes out in a semi-digestible form called a quale that cannot be subdivided even though it contains information on sub-elements of the implicit memory invoked. The reason for this lack of subdivision is simply that the sub-element data is inextricably linked to the quale. Memory however can be broken down into sub-elements by converting it to a place-code address array called a Chunk and editing the Chunk that defines the cluster of mini-column addresses that will be pre-activated. Chunk address lists are probably kept in the Striate cortex where they can be used to influence the mini-column addressing at the thalamus level. This Place-Code addressable form of memory is one I call Explicit Memory.

Declarative Memory

‌The main difference between Explicit and Declarative Memory I suggest is the placement of an index to the memory in the hippocampal area or Meta-Index. This area includes among other things Episodal, Emotive, and Topical Indexes managed by an organ called the Subiculum. The Subiculum also reports to the Striate cortex, allowing it to be addressed in an explicit manner.((indent|8}}Because the Indexes actually store symbolic data about the Cerebral Cortex Contents, the index can be significantly smaller than the Cerebral Cortex, and still reference the contents that are known of the cerebral cortex.

Skill Memory

‌The Skill Memory system controls the sophistication of the physical responses to movement demands placed in the motor areas of the Cerebral Cortex. It involves the loop that connects to the cerebellum, and recent attempts to develop an artificial cerebellum have shown that it might be a useful role to have in a robot.
‌Essentially what it seems to do, is replace a direct movement command with a string of smaller commands that make the movement steadier, and more controlled. There is beginning to be a question as to whether it also acts as a repository for sequences of processing commands, working with the Secondary Motor Area to increase the flexibility and sophistication of both thought and action.

Consolidation

‌Lesion studies and operations on the hippocampal area have shown that it is instrumental in the retrieval of knowledge from the Declarative Memory System. Of especial interest is the Amnesia conditions of patients who have had damage to their hippocampal and medial temporal lobe areas. The fact that these subjects present with both antegrade and retrograde amnesia suggests that there is a 2.5 year window during which a process called Consolidation happens that is important to long-term memory retrieval.
‌While scientists are not yet sure exactly how Consolidation works, they suspect that it is done during sleep, and involves writing the Meta-index to the Cerebral Cortex where it can be used to extract memories. This alternate index, or index image, survives the loss of the hippocampus, and thus allows a certain level of function even after brain damage has resulted in the loss of the hippocampal function.
‌We think that there is a throughput problem here, and the brain must prioritize which memories it updates first, updating the ones most often used before the ones seldom used, which is why students who had a chance to sleep after studying, did better in tests than students who went directly from studying into the test.