User:Graeme E. Smith/Dual Mode Memory

Graeme E. Smith, GreySmith Institute of Advanced Studies

http://en.wikiveristy.org/wiki/Portal:GreySmith_Institute

http://en.wikiversity.org/wiki/User:GreySmith_Institute

grysmith@telus.net

In 1970 David Marr wrote A Theory on cerebral neocortex and described the first 4 layers of the cerebral cortex as a "self-classifying Content Addressable Memory". While some of the details of his theory have proven wrong or misguided by posterity, this remains a cogent description of the cerebral neocortex, and how phenomenally implicit memory becomes accessible. It is by no means a complete description of cortex function, and has been roundly criticized because it makes no attempt to explain certain frequency effects of the natural cortex. Since then, other authors have added to our knowledge of the cerebral cortex, with the result that today our model is much more complete. In this article, I draw from recent work on the Attention System, in an attempt to explain the function of layers of neocortex unexplained by Marr's work, and how those layers contribute to the function of the Cerebral cortex. I draw a parallel between the cerebral cortex, and Dual Port Memory to illustrate my hypothesis.

        Neuroscientists don't like the brain being compared to computer technology, but bear with me I might have a background in computing, but I am not a classical A.I. researcher. Some computer architecture ideas are however, I believe useful as an illustration of a system that might work similar to a neural system, even though the technology (Nerves and stuff) is completely different.
        During this discussion I will use terminology from a number of different disciplines if only because the terminology doesn't translate well between them, because the disciplines are not often brought together in an overview, where the overlap can be seen and the differences counted.One of these terms that is critical to this discussion, is the idea of Demand Memory, Based loosely on Cartesian models {Philosophy} of the mind, the idea of demand memory is that the brain is a reservoir of memory, and we cause it to come to our service, by demanding it. Although most philosophers no longer follow cartesian concepts slavishly, some of the ideas of how memory works, would seem very familiar to a philosopher from DesCartez time. Computer memories for instance demand the contents of a memory area by presenting it's place-code or Address on the buss, and then reading the Databuss to see what the contents were. This should be seen as different from a form of memory where a stimulus, is used by each memory cell to determine if it contains relevant information about that stimulus, in which case it volunteers its contents freely. Such a memory is often called Content addressable memory since it is the content that triggers the output.
          Another of these terms, that I think are useful, is the concept of Dual Ported Memory. In computer hardware, a Dual Ported memory is often installed where two separate systems need to access the same memory. Because each system can access the memory by its own port, it doesn't have to wait while the other system accesses the memory by IT'S port. This article is about a similar type of system that is built out of neurons, and is called the Cerebral Cortex. Where there might be some interference is in the writing of memory, if both systems write to the same memory location at the same time, a collision is said to occur, and one of those systems or both, is going to lose their data.
        Having said all that, I am not someone who believes that the brain is a digital computer. Nor that the analogy is anything but a useful illustration, it can't capture the function of the cerebral cortex completely, because natural systems are often inextricably interwoven with each other, but it does suggest a role for the strange way the cerebral cortex is constructed.
        To look at the Cortex from a perspective that is completely different but comes with illustrations that show details about the arrangement of the cells in the first 4 cortex layers. I like to start with David Marrs A theory on Cerebral Neocortex^[1] In this theory which was an early attempt at developing a theory of the cerebral neocortex, Marr describes a neural circuit he calls a CODON, and shows a schematic of the circuit, to support his concept.
        In this diagram Layer 1 can be seen to be primarily mossy dendrite neurons, Layer 2/3 can be seen to be pyramidal neurons and layer 4 can be seen to be a range of stellate neurons of different types. Connecting the vertical axons of the layer 1 neurons with the Layer 2/3 neurons seems to be some horizontal parallel fiber like transport neurons that carry outputs of the mossy dendrites to the Apical Dendrites of the Pyramidal Neurons.
        What Marr says this circuit is doing is making up a self-classifying content addressable memory element. Theoretically according to information theory, there are two ways that a memory can address it's content. 1 is by searching the content, for the specific memory, and one is to search by a place-code, somewhat like the postal code that the Post Office uses to tell which mailman should get the mail. While some of his assumptions have not held up to the light of more recent work, such as his assumption that the functions of the layer 4 neurons can be captured as basic mathematical calculations, and so the self-classifying element of the system seems more remote than he indicated, I think that the role of the memory as a content addressable memory, can be born out by the neural functions as we understand them today
        Another place that might not survive in today's scientific environment, is his assumptions about Layer 1 sometimes also called Laminae I, which he clearly associates with Mossy Dendrites. Eccles in 1983 published an article^[2] that noted that Laminae I had very few neurons in the Adult Brain. This might however simply be related to the fact that the population of mossy dendrites does not stay constant over the life of the organism. The Mossy Dendrite Layer, is thought to be highly inhibitive, causing the mossy dendrite neurons to compete for output to the rest of the system. The Pyramidal Neurons on the other hand, seem ideally situated to convert the signals from the horizontal parallel fibers, into signals that indicate patterns of activity in the mossy dendrites, and between other pyramidal neurons. It is interesting to note that in Marr's many models of Cortex Tissues, almost every case associates pyramidal cells with close proximity to Mossy Dendrite Neurons. The one possible exception being the hippocampus or Archeocortex, where the Mossy Dendrite Neurons are in the Dentate Gyrus rather than in the hippocampus itself. It is the feedback from other pyramidal neurons that best supports the idea of a self classifying system. Essentially the pyramids are the output neurons for the content addressable memory, and their distinctive pyramidal shape might have to do with the basal dendrites that interconnect them.
        Layer 4 neurons can be seen to be both a special input layer for Thalamo-cortical fibers that have the senses as their source, and as a management layer, that allows other brain systems to reach into the context addressable memory and influence it in a number of ways. Marr of course thought that these cells did mathematical functions on the pyramidal outputs, but later work has thrown some discussion into the mix, by pointing out that the nature of the connections is more random than Marr would have thought, and so mathematical functions are less likely to be terribly effective, where process management functions might be much more effective.
        The difference is less one of Function, as much as degree, Marr thought that his Probabilistic Functions would give him the ability to detect patterns in the connections between neurons that never materialized. His failure brought home to neuroscientists just how individualistic brains were, and eliminated the idea of the circuits being comparable mathematical functions.
        Despite Marr's failure to explain the connections, and his eventual death in 1980, his theory remains a landmark in Neuroscience because it approximates the function of the first 4 layers of the cerebral cortex. There is still some question as to whether the cerebral neocortex is actually a homogeneous function or whether there are areas that perform other functions, but it is a useful theory if only because it opens the question of content addressability in Neural Systems.
        Before I go on, I would like to introduce to you, the concept of cortex micro-anatomy. Essentially the Neocortex is primarily a 6 layer cortex, numbered inwardly from the surface of the cortex in layers or laminae. Layer 1 or Laminae I is the outer cortex, and layer 6 or Laminae VI is the inner layer of the cortex. Such 6 layer cortex tissues have been called Isocortical tissues or Granular tissues. A variation of this architecture has a missing granular or laminae iv layer, and is therefore called Agranular. Other tissues have 4 or less layers and are called Allocortical tissues. This should be kept separate in your mind from Allocortex Tissues which are tissues in a specific part of the cortex, (The Allocortex) not a type of tissue designated by its micro-architecture.
        It can be said that most of the tissues in the Allocortex are Allocortical, and perhaps most of the Allocortical Tissues can be found in the Allocortex, but some Allocortical tissues are found outside the Allocortex. In fact it has been noted that the tissues underlying the Sulcys, and Divides of the brain are Allocortical and they exist in nearly all parts of the Cerebral cortex. Some of the Allocortical Tissues can even be found in the Archeocortex, but they are a special case because they are not stacked vertically like other cortex elements but tend to wrap or coil around each other. If we look closely at Marr's model we see that he is actually describing Allocortical Tissue not Isocortical or Agranular Tissues. However, there is no reason to think that Isocortical Tissues or Agranular tissues are differently organized in their top 4 layers, there are just 2 other layers missing from Marrs model. If this is correct then maybe just maybe Allocortical Tissue is content addressable memory, and Isocortical Tissues have some added function.
        The next thinker I would like to introduce you to, is Jerry Fodor. Fodor is well known for his Modular Mind theory, but it is his work in the representational limits of Computational Psychology, that I want to draw to your attention. In The Mind Doesn't Work That Way!: The Scope and limits of Computational Psychology^[3] Fodor's comments lightly touch on the fact that Phenomenal Systems like Neural Networks are unsuited for isolation of specific memories, because of the distributed nature of the processing and storage functions.
        In essence what he is telling us, is that the representation limits the function of memory in phenomenal (Neural Network) systems. Now this is a much more important concept than I think even Jerry Fodor knew when he wrote about it, He probably wouldn't have passed over it, as quickly as he did, if he knew the value of the concept, but his main push is Modularity, so he filled the remainder of the book with representational stuff about modularity.
        Now we get into a philosophical concept called "Phenomenality" Phenomenality means that much of our experience can't be broken down into smaller parts. Along with this concept comes the idea of a Quale or indigestible lump of data that refuses to split into smaller parts. Phenomenalist Philosophers are always going on about Qualia, and how we can't represent the color red, even though everybody understands what it is. The it, that we understand may not be at all the same as the It that everybody else understands (Especially if your are Red-Green colorblind). There is quite a spirited discussion going on on PhilPapers (http://www.philpapers.org) in a thread called the "Explanatory Gap" where the Phenomenalists are arguing with the physicalists as to whether there is some part of our experience that we will never be able to explain. This something we will never explain (some claim)is called the explanatory gap. The fact that we have not yet managed to explain it, is held up as at least an indicator that it is a "Hard Problem" in Consciousness Philosophy.
        So why would something exist that couldn't be broken down into smaller parts? The answer lies in a not well understood aspect of Content Addressable Memory. We are more familiar with demand type memories, in fact we assume that our Declarative Memory is just such a thing, we demand a memory and it appears as if by magic in our mind. But in a Content addressable memory it is the Content of the stimulus, that triggers the memory not the demand for it. And, as some scientists have learned to their sorrow, If once you can do something, someone else will be able to do it too. David Marr noticed this in his article back in 1970, he thought that most of the processing in the brain was probably being done to deal with redundancy. If we see content addressable memory as places with contents that trigger at the hint of anything that looks like them, then we can envision a flurry of contents being released in a big data-cloud all at once. Now add to that flurry of activity the idea that at the neural level every brain is individualistically connected, in other words no two brains are alike, and the idea that the location in which a memory is placed is arbitrary to the network and its state just before the memory was placed there, and you get a big ball of data, that can't be sorted directly. In other words you get a Quale. So one thing we can say about allocortical memory is that it is phenomenal and it's output is a Quale.
        What this suggests is that Demand Memory, the ability to isolate specific memories requires a redescription to a different representation in order to work. This is counter intuitive to computer guys like me, because computer memories are so easy to isolate that there is a tendency to think in terms of remapping instead of redescription. This is also where the Dual ported memory seems to break down, in that if we have two different ways of addressing the same memory, can we say that the memory has been redescribed? Unless, of course as in this case, it is the addressing mechanism that makes up the new representation and conversion is mostly a case of trading one addressing technique for another.
        In this case, in fact, I propose a hypothesis that it IS the addressing technique that is being redescribed, not the content of the memory. This brings us back into frame with the Dual Ported Memory illustration. It makes sense that if two different systems use two different addressing techniques to access the same memories, that in a dual ported memory, the main difference between the two memory ports would be the techniques that they use to access the memory. Now suppose that the two addressing techniques are mutually exclusive, that you can't transfer information between them except through the dual ported memory. In this case you have a problem, either the data has to be formatted in the originating system to be retrieved by the using system, or the using system has to have some way of searching the memory in order to find out what it contains. In computer programming this wouldn't be a problem as long as there was documentation, a standard by which we could translate memories into a form useable in the other system. In fact a lot of computing is simply finding the right symbolic standard for transferring information from one format to another
        Natural Neural Networks however, do not come with manuals, and they are phenomenal, in the sense that the location of a memory is a function of the whole network, and it's experience. In other words location of memory does not have a standard. So if you don't have a standard storage format, you can't have a standard translation package waiting ready to go. Now add the fact that your data is all balled up in a Quale, and it's enough to have Dr. Fodor asking whether it is possible to implement a phenomenal demand memory.
        So this is the problem that requires redescription. Phenomenally Implicit (Content Addressable) memory can't be translated into demand memory, without some form of search during which the memories are redescribed into a demand memory addressing format. Having described the necessity for that role, I will mention that redescription is not my own idea, but was explored in Annette Karmiloff-Smith's book, Beyond Modularity: A Developmental Perspective on Cognitive Science^[4] . I will leave the discussion of how the redescription is achieved for a different article, and concentrate mostly in this article on the nature of the dual ported memory that it implies.
        There are two problems with implementing a Demand type memory on neural networks, the first is that neural networks tend not to have stable storage locations, so place-code addressing them is difficult, the second is that you would need some method of addressing the stable storage locations if they did, but the location of any particular storage element is indeterminate in natural neural networks.
        In order to understand how the brain could possibly stabilize neural networks to have at least semi-stable storage locations we need to understand the idea of Neural Groups. A neural group is a cluster of some hundreds of neurons that acts as a single unit when it comes to firing. It is how we get to this point that is tricky, and the research is still ongoing to fully explain it. What we think is happening is that the pyramidal neurons in the cortex, are attached by basal dendrites that sense the pre-firing activation levels of their surrounding neighbors. This is a little controversial because it was thought that firing was the important part of neural function at one point. One way of looking at this, is as a community of neurons all encouraging each other to fire.The more neurons in the group, that encourage firing, the more likely that a neuron in the group will fire.
        Now we add to this the idea of a Center Surround, in that if the group fires it is usually one or two neurons that fire, and the rest stay dormant. It is as if there were a suppressive field that acted to inhibit the surrounding neurons and just let the centroid of the set fire. We think that this effect is mediated by the 6th layer pyramidal neurons, which form distinct columns of neurons that fire if the column is stimulated at any layer. What might be happening is that the 6th layer pyramidal neuron is regulating the suppression field needed for the center surround function. A candidate for the generation of the suppression field, is the Martinotti Neurons which act on layer 1 to suppress the firing of layer 2/3 neurons. If the sixth layer pyramidal neuron fires, so does the column, and by extension the Neural group it contains. So the sixth layer pyramidal Neuron balances the inhibitory effects of the Martinotti Cells. By allowing only a few neurons to fire at the same time.
        Now why does the brain work in this manner? the current concept is that the columns reduce the resolution of the memory, causing it to cluster like signals together into a single signal. This might have the added effect of causing the the memory to stabilize the location of a memory when it tends to reinforce other memories around it. Sort of helping the self-organization of the memory, so that it is also self-classifying. If this latter interpretation is true, then it suggests that in fact the existence of Neural Groups might promote the stability of the memory locations.
        To get at the second requirement some way of addressing the stable neural groups we need to step outside the cerebral cortex, and look at the attention system. Dr. David LaBerge has worked on the Attention system and has written several articles on it, that might be useful. the first, is an article called Attention, Awareness and the Triangular Circuit of Attention^[5] a second is Attentional Control: Brief and Prolonged ^[6] Together what these articles suggest is that there is a form of bottom-up attention mediated by the thalamus that connects to the 5th layer apical dendrites acting to pre-activate clusters of neurons that he calls mini-columns which are part of the larger columns.
        What the triangular theory of attention is about is the fact that in order for a person to become aware of something, three areas of the brain need to light up simultaneously. The Cerebral Cortex, where the memory will be stored, the bottom-up attention system as moderated by the thalamus, and the Top-Down Attention system which is moderated by the PreFrontal Cortex (PFC). What the Attentional Control: brief and prolonged article is about is how the cortico-thalamic connections use the 5th layer pyramidal neurons to trigger the pre-activation of 2/3 layer pyramidal neurons in small neural groups that in combination make up the larger column. Since pre-activation is the mode in which neural groups are triggered this pre-activation step is all that is needed, in order to address a memory by triggering a neural group to fire.
        In other words, this 6 layer model of the cerebral neocortex, (An Isocortical Model) includes everything we need to have a phenomenally implicit content addressable memory, and a demand based portal to the same memories that were accessed formerly solely by content. It defines if you will the Dual Port Memory, that is needed to achieve both addressability by content and by demand. However be advised I have intentionally glossed over how the information is redescribed so that the demand based memories have meaning. I have just pointed out that it must involve a search, because the two types of memories are incompatible in a neural network system, and no reformatting by mapping is possible because no map is possible at the connection level. Thus the 6 layer (Isocortical) architecture that makes up the Cerebral Neocortex, describes a mechanism whereby implicit memories can be retrieved using a place-code addressing scheme based on mini-column pre-activation. There is only one problem with this theory, and that is that the Quale that is output by this addressing, is still an indigestible data cloud with redundancy and a real lack of order. I will endeavor to clear this problem up later in another article by describing the stages of redescription necessary to create accessible memories.