Brain Physiology, Cognition and Consciousness group: topic
This is a public discussion board
From Neuron to Memory System: How Memory Might Work
Graeme Smith
Friday, 12 June 2009 04:34 UTC
How does memory work?
We think we know, based on thousands of years of research into it, but there are still indications that we might be completely wrong. Scientists created computers with the best scientific approach to memory based on current scientific theories, in the 1940’s, and they got it wrong. Most neuroscientists today will tell you that the computer is nowhere close to a good model of memory, yet many of the attitudes that informed the decisions on how to make computer memories remain unchallenged. Our popular theories of how memory works were laid down in the middle ages, by people who thought memory was a fluid, and that the brain was a reservoir. Somehow they thought, we opened valves, and let the memory flow from the brain to where it was needed. I call this a demand model of memory because you demand the memory by opening the valves. In computing terms we have replaced the fluid theory of electricity but retained the idea of current, and at one time tubes which were replaced with transistors were called Valves. So computer memories could be seen to be operating in a demand memory model as well.
As Jerry A. Fodor said in The Mind Doesn’t Work That Way!:The scope and limitations of Computational Psychology Neither Phenomenal approaches such as Neural Networks nor strict computational approaches such as Truth Preserving Functions, seem likely to achieve a suitable computational model of consciousness.
As a researcher in Artificial Consciousness, my main thrust is to eventually get support for my own Artificial Consciousness to be built. However to get there, I had to start with a model of memory. I chose to hedge my bets, to begin with a Neural Model, and add Functional support to that model, where it was needed. Further I wasn’t going to limit myself to Truth Preserving Functions, where Soft computing would be more effective. This type of approach to consciousness is called a Hybrid approach. To start off, however I wanted a model of memory, and one of the things I had to do, to get there was overcome Fodors assurance that no-one had yet developed a phenomenal version of a demand memory.
Those who are stuck on William James idea that phenomenal systems can’t be built based on what we know about the brain, have problems with my use of the word phenomenal in the way that Jerry Fodor used it in his book. They like to look for exotic things like QM entanglement to explain the fact that some things just seem indivisible. I however think that there are indivisible elements in the memory but that it doesn’t matter because the mind doesn’t try to divide them, it finesses the system, without that effect.
Fodor’s claim is not without merit, phenomenal systems are definitely easier to build out of neural networks, than demand memories. However, David Marr defined a type of memory he thought might help explain the cerebral neocortex back in 1970, and by understanding what his work uncovered, and brushing off some of the unfortunate assumptions of his day, I think I have undercovered one of the hidden biases that is keeping us from understanding the way the brain, or at least the memory works. And that bias is the assumption that explicit (demand Model) memory, is the natural state of the memory system, and that anything that doesn’t fit the model must be an add on to the basic explicit memory model. So we expect Implicit Memory to somehow be an add-on to the essential demand memory, and we are wrong.
What Marr described in 1970 is a type of memory more basic than explicit memory when implemented in neurons. To computer guys like me, this seems counterintuitive. How can anything be more basic than a good demand memory, we can implement dynamic ram with a capacitor, a resistor, and a transistor, how much more basic can you get? But what we keep forgetting is that neurons don’t work like transistors, there is no technological overlap. Within the logic that makes neural networks work, demand memory is much more expensive to build than a simple content addressable memory. That is what Marr claimed his 4 layer CODON was, was a self-classifying content addressable memory.
To understand why this might be, we need a little theory. Although in the 80’s and 90’s the connectionist school was over-run by the Parallel Distributed Processing guys from the A.I. discipline, and so we have to take some of its theory as being deliberately misleading to steer people away from trying to understand real neural systems, The basics are fairly equivalent. A neuron is first of all a cell, it has to survive like other cells, by absorbing nutrients and getting rid of wastes. However at some point in the evolution of animals, neurons gave up some of their survival functions to a helper cell called a glial cell, and converted those functions over into mechanisms for transferring information between cells.
The Parallel Distributed Processing guys figured what was important to this communication was the firing of the cell. Sorry that is a misconception, firing just speeds the process of transferring information up, it is not the only mode of transfer, nor is it the most important one for understanding natural systems based on nerve cells. However because the PDP boys wrote the manuals for the industry, they got to tell 20 years of modelers what to think. As a result of mistakes like this, they set the neural modeling of Natural Neural Networks back, to the point where Gary Marcus, has clearly stated that he feels it necessary to do a hostile takeover of the connectionist school.
One approach that the PDP guys thought was dangerous was David Marr’s attempt to use probability Mathematics to define a circuit that was made up of heterogeneous neurons. Despite the fact that they did not have the ability to model heterogeneous groups of neurons, a direct attack on his claiming that the heterogeneous group he called a codon was a content addressable memory, was made pointing out that the model did not exibit wave-forms similar to the real cortex.
This attack, assumed that Marr’s circuit had to have frequency artifacts similar to the real cortex in order to suggest the role of content addressable memory for the cerebral neocortex. Well Marr was a pioneer in the field and he got some things wrong, and others even more wrong, but it was a seminal stage in the science, and what he got right is more important.
Marr was able to use probability mathematics to analyze heterogeneous networks of neurons and predict their function. Nobody before that even tried, and after the way he was treated few after him had the nerve to try again.
I am not going to start spouting probability equations, if only because I don’t follow his math. But the PDP guys called a neuron a processor that had input, process and output capabilities. I however look at it from a different perspective I see a neuron that has storage, processing, and transfer capabilities. And when I look at the 4 layer cortex model that Marr worked with, I also see a content addressable memory, but I think that perhaps he expected too much equivalency between neurons in the connections department, and was sadly dissapointed before his death in 1980. So I wonder at how much self-classification the system is doing.
If you look at neurons from my viewpoint you begin to see that the shape of the neuron has a logic, that memory neurons have lots of synapses, and opportunistic process growth, that processing neurons tend to have complex and often bushy or mossy dendrites, and that transport neurons tend to have virtually no division at the dendrite level and form long thin neurons.
If you look at the type of organs that are produced, Memory Neurons tend to form tissues, Processing Neurons tend to form globular organs, and transport neurons tend to form fibrous bundles. In other words the shape of the neuron, indicates it’s function, and the shape of the organ might indicate to some small extent what type of function it performs.
Looking at Marr’s 4 layer model, we see that the first layer is processing neurons, the second and third layer are memory neurons, and the fourth layer is dominated by processing neurons again. The whole structure is probably a memory role because it is a tissue.
Stimuli coming in at the first layer, are processed affecting their storage in the second and third layers, and the 4th layer probably is mostly control neurons, and in some cases inputs of senses that have been directed through the thalamus. These latter signals condition the general inhibitive environment caused by the Martinotti cells and thus reduce the inhibition at the first layer encouraging the layer 2/3 neurons to fire.
This simple memory circuit can be seen to be designed to respond to either signals from layer 4 or layer 1, and the first layer is used to train the second and third layers to respond to patterns of stimulus. Given a stimulus each pyramidal cell in layer 2/3 decides whether or not to fire, and this type of system is exactly what is needed for a content addressable memory.
It is also exactly wrong for an explicit memory suggesting that it must be the explicit memory that is the add on. Indeed when we look at the micro-architecture we see that a type of tissue called Allocortical tissue makes up the tissues at the bottom of the Sulcys and Divides of the brain. and since Sulcys and Divides are a later architectural configuration there might be an indication that the six layer tissue called isocortical tissue, that pushes the gyruses up away from the sulcys and divides, was a specialization of the brain.
-
Replies
Jump to resultsResults
-
If we continue with the model as it has been presented so far, for every context that we have previous experience with to the point where the traces are set into the purkinje cells in the cerebellum, there are a number of processes that could be selected. Given that each separate process is given its own ASO tag, the ACC can select one of the processes simply by suppressing all the other ones. The problem is to know which one to select by suppression of all the other options.
One way of selecting the next step in the process, is to predict what the result of taking that step might be for each option. The best way of doing that, is to run a model of the step, based on the best estimate of how it will react to the world as we have percieved it, and test that model against a set of limits that define a “Comfort Zone” within which our bodies are best run.
Another way of looking at it, is we run a model of the action/process against a model of our self, to determine if the action/process will push us out of our comfort zone.
Now here we get into hypotheticals again, It is my contention that the set-points for the self-model are set during developmental sweet spots, and act as presets for limiting behavior. Certain psychological conditions might exist as a result of the failure of the developmental process to achieve these preset conditions, and thus the failure of the model to adequately limit behavior. It is possible for the self-model to adapt to such a failure of development, however the result is an adjustment point which is not stable and can be raised or lowered according to limbic function. People dependent on these adjustment points, would tend to manage to stay in acceptable ranges most of the time except when they were in situation of high emotional content, at which point, their adjustments would slip and they would step outside acceptable bounds for behavior.
Here we get into the concept of Agency. An agent is a self-guided process, along with the self-guidance that such a model/self system would offer, comes the sense of self, a meta-cognitive “Feeling” my hypothesis is that the feeling of self, is meant as a signal that a process has passed the modeling test, and an affirmation that we evaluated that it would keep us in our comfort zone.
-
The model of Agency I use is a three part model, consisting of:
1. the model/self test for range violations
2. the Meta-Cognitive signal that the test has been passed
3. A Log of activity done while the meta-cognitive signal is active.
This seems to cover most of the modes of errors in agency found in pathology, although I am by no means expert on the topic.
The question becomes where did the log come from?
The answer comes from computer language theory, and the idea of a rewindable macro. Essentially any process can complete in an abnormal condition, or failure mode. This is especially true if the parameters of the process are changed between the time it is laid in, and the time it is executed, say by optimizing it or something.
In an adaptable Macro language, both the concept of optimization and the concept of rewinding a macro if it fails, make good sense. For instance, part of a macro, is the interpretive overhead of selection and execution of the macro itself. During compilation the removal of this overhead, would act to speed the macro, making it more optimal. However, in some cases taking this overhead out would be an error, if only because a parameter needs to be set, and it is only during the calling that the parameter is referenced.
Any automated system of optimization is therefore going to result in a failure a significant amount of the time. However Optimizing compilers have proven that as long as the percentage of failure is low, it is actually better to optimize and fail and have to rewind the process, than to ignore the chance to optimize. To rewind a process, you need a log of the call of that process, or somewhere to store the call, (such as a stack) while it is being executed. It is my hypothesis that lacking a stack, humans make use of a log, creating a pseudo-sequence of macro-calls that can be used to rewind a process.
For reasons that might not be immediately obvious, I call this log sequence, Awareness, and suggest that because of the need to decide whether or not to rewind, this sequence is analyzed immediately after the actions it reflects.
One of the things it does is check to see if the results of the processing stayed within bounds, by re-applying the model of self to them. If they didn’t then a fault is triggered, and a test is run to see if the process can be rewound.
-
You may wonder why I am talking about intention and awareness, and agency in a thread on memory. One good reason, is that Awareness in this model can be seen as a sequential buffer, that always follows what just happened a few milliseconds ago.
The present, is remembered right after it occurs. I am not sure where in the brain the buffer for awareness is, but I suspect by the type of information that it gathers, that it might be in the prefrontal cortex somewhere. It is my contention that the meta-cognitive signal that imputes self-hood to the log, comes after the laying down of the log, and therefore while associated with the log, represents part of further processing not a component of the log.
Thus I suggest that awareness does not carry the meta-cognitive signal of self-hood. If there is a reflective system that does reflect the meta-cognitive signal of self-hood, associated with awareness, as Dr. LaBerge claims, then that reflective log, would probably be consciousness.
Since it is often difficult in a reflection of a reflection to see where one reflection ends and the later reflection begins, the cognitive reflective logs are difficult to separate, and may be mistaken for a single log. This implies a self-referential state where the output from the reflection feeds back into its own past to allow meta-cognitive signals to co-locate in time the logs they are analysis of. Much better would be to recognize the time lag between an event and conscious reaction to it. This seems to support Libetts work, where Volition doesn’t start until 500 milliseconds after intention does.
One reason that this does not feel like a practical interval, is that our sense of time is associated with conscious mirroring not intentive mirroring, and therefore we don’t experience the sense of passing time while doing intentive tasks. From the point of our internal mental clock, time is stopped until the conscious reflection is processed. This means it is possible to lose all track of time, while operating at a high skill level for an extended period. In my own case, I can read a book, without a sense of the passage of time, unless it requires conscious reflection. One reason I do not do as well with text books as with science fiction is simply that I have to think more about text books than I do about science fiction. This requires conscious reflection, which breaks my concentration, and fatigues my brain more than reading a science fiction novel does.
-
Dear Graeme,
I do not see how awareness can be conceived as a buffer of some kind. In that model only the content is addressed. What seems to be missing is that the content is witnessed, is experienced consciously, whatever you want to call that.
In another sense you are very close to the views of Global Workspace Theory.
Also a caveat Alfredo and I expressed in our recent article: the present can be remembered, but must not be remembered. Remembrance is not certain. Memorizing and memory retrieval follow certain rules, but we are far from knowing which those rules are and what the factors are (factors are probably many and will change all the time) that influence this process. We know part of it for sure.
You wrote:“Much better would be to recognize the time lag between an event and conscious reaction to it.”
That is so at times and not so in many other occasions. I do not even think it is a good idea to speak of a conscious reaction. Reactions I would rather define as unconscious. If a conscious response happens there does not have to be any given time lag (or gap to be measured in milliseconds) – a conscious response can happen whenever the individual consciously choses that this response is the one that fits into the conscious evaluation of that individual.
Conscious evaluation may take more or less time depending on the circumstances.
Yours friendly
Hans -
Thanks for your comments Hans, I knew I was going to get in trouble with the consciousness mavins when I linked consciousness to a buffer. That is why I have stayed away from the Consciousness threads, I really didn’t want to get into an argument until I had laid the groundwork, in the rest of my thread.
There are a number of differences between my work and Global Workspace Theory, first of all, I believe in a limited STS, but not that it is the NETWORK that limits it. Instead my theory stresses the conversion between implicit and explicit memory as the choke point.
Secondly, there are a few more steps in between before we reach the Global Network, and I have limited the network to a global connection to the ACC and a more limited set of connections via the Commissures.
Thirdly there is the mechanism by which functions are selected, which moves away from GWT’s multi-agent architecture, such as implemented in IDA, and towards selection/filtering via the ACC. Secondly GWT does not deal with the Core/Belt/Associative area architecture at all, unless I have missed an update or two.
But thank you for putting my work even on the same page as Bernards.
On memory retrieval as associated with consciousness, I have to say, that you must have missed a point, and it is probably my fault if you did, seeing that I am having so much trouble expressing the model.
Awareness and Consciousness in this model reflect the macro-language execution stream, not the main memory. The macro-language of course contains references to data elements that have already been chunked, so that all that is required is that the chunks be rehearsed to inform consciousness as to what the data was. This might be why the confusion between remembering the present, and retrieving memories of the present might be found.
Libett’s work suggests that consciousness takes processing at the intentional level to achieve. So your assumption that consciousness does not have a lag, may be because you are unwilling to part with the idea that consciousness is the causative factor. If consciousness is instead a phase of processing that comes as a result of the failure of intention to resolve choices for the ACC, then the lag is required in order to qualify for its activation.
I hesitate to turn this discussion into a discussion on consciousness, since I am not sure we are using anything like the same term. I was staying away from the consciousness discussions for that same reason. If you want to discuss this model of consciousness I really should start a new thread based on my model, so I can fully introduce it to you, instead of starting just with those aspects of it, that are related to memory.
-
There is a new thread called From Memory System to Mind, How Mind Might Work where we can discuss the consciousness model.
-
I have not shut off this thread yet, if only because I think that the principles and insights that make it work are difficult to think your way around if you come to it from a classical psychology/neuroscience/A.I./Philosophy background. That means that most of the readers will probably still have questions once they absorb the model a little more, or will have already thrown the baby out with the bathwater so to speak, and have rejected the model out of hand.
This is a valid approach, since I am not kowtowing to current wisdom, nor letting myself be deflected by distractions. As much as possible I am integrating the data behind the wisdom into my model where possible, and this will be seen as a rejection of wisdom, which is not usually seen as being wise.
I have never said that this is much more than a model of How memory might work, that is part of an Artificial Consciousness Project. As a model, there are inaccuracies mostly of the nature of over-simplistic explanations, and there are places where the model might not fit the reality of the brain.
Models are like that.
What is important is how the brain looks through this model, as an integrated Cognitive Architecture mostly based around the requirements of memory.
This model does something that no other model of the brain has ever done to my knowledge and that is draw a practical connection from the neuron to the psychological models of memory, through each intervening layer and try to explain how those layers work together to form the outputs we can detect with tests.
It is by no means complete, and may be out of date in many ways since the research it is based on has been published and so by definition is out of date.
But it offers a glimpse of never-before seen compromises that are critical to brain function, and begs the formation of new terms or changes in the use of terms to better describe the systems involved. If there is one thing wrong with it, it is that it is too precocious, and claims to explain too much, but if you can follow the rationale, you will see that it has a reason although by no means any proof yet, for making those claims.
The reason I am publishing it here, is simply that I wish to learn how to express my ideas better before I try to capture the attention of those who claim to be experts in the field. This might seem upside down to those of you that have the benefit of a degree in science, but I have no such benefit, and must use the tools that I have available to me.
This model is based on the idea of a complex neuron capable of learning both short term memory and long-term memory within the same cell. This flies in the face of much historical work aimed at finding out what parts of the cortex were specialized for long-term memory. It is not that there aren’t such areas, but that they have roles in different aspects of memory than we originally thought. It also asks for a redefinition of how we think of Short Term Memory, which is very heavily influenced historically by the work of Miller et al. Part of the problem is that we are using the same term for two different types of memory, begging a redefinition of one or the other to a new term. I originally tried to define such a term, but got in trouble with pre-utilization of the term to mean something else. Learning my lesson, I leave redefinitions to Academics that have the time to search through lots and lots of articles to find out what terms are already in use.
Even when I try to make use of the definitions as I find them in particular articles or books, I get in trouble, because quite often different disciplines or schools redefine the same terms to mean different things. One such term argued over in this thread is the term Phenomenal. I used the term two different ways, and it made sense to me, but those invested in the terms quickly objected to my use of both of them. C’est La Vie, I am not an academic to argue terms. I could certainly care less as long as the idea of how I meant them to be interpreted gets across.
Above the neuron lies the concept of Heterogeneous Networks. While Neural Computation models some of the characteristics of different types of neurons, they tend to do models using fully connected networks. One aspect of heterogeneous networks is the requirement of the model of the neuron in the network to allow for a more complex connection scheme. These networks can be seen to be built out of Groups of different types of neurons which is what I call a Heterogeneous Group. These heterogeneous groups are often at least in the cerebral cortex, found in arrangements that have a noted laminar character when stained in various stains. Interaction between the layers is a critical part of understanding the mechanism of how Mature Isocortical Tissue works.
Along with the Mature Isocortical Tissue, there are areas of less mature, cortical tissues, and areas that have fewer than the requisite six layers that make up an isocortical tissue. These areas are called Allocortical and a great many of them lie within the Allocortex, or areas surrounding the hippocampus. The hippocampus itself can be seen as an early form of Allocortex that has not yet formed a tissue structure.
It is my belief that the Superficial Layers of the Isocortex are really a three layer Allocortical arrangement that has the role of acting as an implicit memory. The remaining layers consist of layer IV which most often is involved in redirection of senses via the thalamus, and Layer V, VI which are, I suggest involved in the stabilization and addressing at the mini-colum/column level. Mini-columns seem to have the same approximate resolution as Neural Groups, which have been noted to fire as a group rather than firing as individual neurons, so I suggest that there is something in the mechanism of Layer V, VI that implements the neural group arrangement.
While by no means proving this contention, there is support found in the Architectonics of the layers, at the Core/Belt/Associative layer where fibers connecting the core to the belt, and differences in staining between the core and belt areas, suggests transport of information in the direction of core to belt, rather than the other way around. This area called IIIc seems to be added onto the basic implicit memory possibly as a way of including data from the core, in such a manner as to recognize patterns at the belt level.
A parallel is drawn between the core/belt/association areas and the three types of explicit attention suggesting that in fact the implicit memory in the belt and association areas is influenced heavily by explicit addressing of data from the primary or core areas. A possible use for the belt is projected as being a place where relationships between zones found in the core is developed to form images of sub-zones within the perception focus that might correspond to separate memory elements.
The explicit addressing of data in the belt area, is often mixed with the complicit attention mechanism which is both an explicit addressing and functional addressing scheme that is taken to implement a command language of undetermined nature. It is thought that the modules in the Association areas are triggered by preactivation similar to that used to define the data in the other two areas. The net effect is that attention on the belt area, and attention on the Association Area, triggers a set function or command that automatically processes the belt data through the Association function at that particular module.
In some cases a case can be made for recruitment outside the Association areas, this might be due to the mixing of sensory modalities at the borders of the Association areas, or it might be due to a failure of classification of Association areas outside the Localized structure of core/belt/association areas.
In order to make this memory/processing system work, however, there needs to be some sort of an index by which specific elements in the Belt Areas can be addressed. Since the belt areas consist to some extent of data derived from the core areas, the actual elements found in them, are dependent on data throughput, not just the arrangement of connections. Furthermore, Fodor assures us that addressing in an implicit memory is not trivial, and even though we have an system mapped at the neural group level, we will not be able to map that data on any rational map because of the nature of neural networks. Dr. Edelman concurrs saying that neural groups are interchangeable, and therefore the addressing code needed to address them is degenerate (having multiple terms for the same data element) and redundant.
For this reason finding memory elements is difficult not because we can’t address the neural groups but because one neural group does not map out to a single storage element, nor is it practical to expect a common function to have evolved across any single species, that speeds mapping. Instead we have to accept the need for an index.
The Brain seems to solve this problem by utilizing certain semi-mature cortex structures in and around the hippocampus. For the purposes of this model I call these the Meta-index. One aspect of this meta-index is the Episodic Memory that has been found to exist in the hippocampus. Another aspect might be an immature cortex structure that has 4 layers, and a notable mottled appearance found in the Entorhinal Cortex. This area might be an index for finding concepts in the Cerebral cortex.
The cerebral-hippocampal loop probably describes the brain areas that are involved in implementing the meta-index. The reason that it is a loop, is that the locations of cerebral data are encoded in the meta-index, and then the meta-index is used to access the cerebral data elements.
There are three such loops in the brain, and the third loop, connects the brain stem and cerebellum with the cortex. Confusion happens because there is a secondary circuit through the thalamus, which is part of the main memory loop being the preactivation center for the cerebral cortex as well as a router re-routing data from the senses through different circuits in the cerebral cortex.
It is thought that the brain-stem loop might be the source of skill memory, and it has been noted that damage to the Meta-index loop does not affect either implicit or skill memory, which the parallel nature of the circuits would suggest is the case. However damage to the Meta-index does affect the retrieval of Declarative memory, and may influence long-term memory.
This certainly fits within the index model, especially if the type of long-term memory involved might be an archive of the Meta-index updated periodically according to time restraints, from the image currently within the Meta-index over a period of up to 2.5 years. Because the damage to the medial temporal cortex, and hippocampus involved in H.M.’s case almost completely removed the hippocampus, the location of the index archive image may be somewhere in the cerebral cortex, which would explain why some areas of the cerebral cortex are thought to be associated with long-term memory.
A model of Cerebellum function developed in the 60’s suggests that the role of the cerebellum is as a sequence memory for actions. I have chosen to suggest that this might be expanded to include sequences of functions as well, creating a Macro-Language.
The role of the cerebellum in serving up alternate activities via preactivation via the thalamus, suggests a role for the ACC in selection from among equivalent actions/processes. This links the skill memory to the Attention system at both top-down and bottom-up layers, and presents an opportunity for programs to be built, and stored and reutilized.
Linking the role of the cerebellum with PFC areas that are linked in turn to the Secondary Motor Area, with its tendency to gather sequences, and serve them up at a higher level, I draw a parallel to the Forth Languages minimalist two loop interpreter. The cerebellum acting as the inner loop that follows the sequence of each separate macro, and the SMA acting as the outer loop that follows the sequence of macros. I also implicate the SMA in the selection of models, possibly through passing the parameters of the macro into a modelling mechanism based on the inferior hemispheres operation and blocking, or redirecting of the actual activation so as to limit interference until the model has been tested.
In any case, the results are a heirarchy of control mechanisms. The basic blind attention system of a naive individual based on random impulse, the more advanced self-programming system based on skill memory, called intention and an even more advanced level of self-programming, called volition that carries with it the functions of Consciousness, at least at a minimalist level associated with animals.
These last two layers because of their self-programming nature need to be regulated to keep the programs within the comfort zone of the individual organism, and as a result, require feedback mechanisms that are defined as the Awareness Buffer, and the Consciusness Buffer.
That more or less captures the Recap of the nature of my Memory System Model.
In any case attention being again a primary part of the processing.
Results
-
