Artificial Consciousness/Neural Correlates/Functional Models/Memory
I have already more or less Introduced the concept of memory in my article on Attention. Memory comes in at least six types as follows:
- Conditioned or Learned Reflex
- Implicit Memory
- Explicit Memory and Working Memory
- Declaratvie Memory
- Episodic Memory
- Skill Memory
It is interesting to note that most of these types of memories are related to a Homogeneous Neural Group of Pyramidal Neurons. This is probably because of the nature of long-term memory and the demands it places on the cells that perform it. It's not that the whole brain isn't made up of cells who have memory, it is just that Pyramidal Cells are better placed for a specialized role as memory elements. No memory cell performs pure memory functions, they all do three things, store, process, and transport data. It is just that some cells are more specialized to store and retrieve data than others. The Pyramidal Cells don't work alone, they often have a relatively large number of support cells working with them. A Heterogeneous Group based on the Homogeneous group.
Memory starts at the Neuron Level at the Synapse, where connections with other neurons are formalized.
We will stay away from over-used labels such as short and long-term memory and just mention that Synaptic memory changes naturally over a relatively short period of time, As signal flow ebbs and flows throughout the brain. One limiting factor is the fact that the synapse is dependent on the sensitive patch on the cell membrane, and natural digestion processes within the cell would denature all its proteins. In order to keep this from happening periodically the proteins are taken out and replaced by the membrane replacement mechanism. As part of this process, some of the proteins that have not yet been digested, are taken out of the old membrane and reused later. As a result of this storage, they take longer to digest, but digestion still occurs over time. When digestion gets far enough along, the proteins quit doing their job, and get dropped from the sensitive patch during replacement. Since the active element in storage at this level is the Ion channel that sets the weight of the synapse, Synapses that are stronger connections collect more Ion channels, and signals that are weaker connections gradually lose ion channels as the aren't replaced. We think that the main difference at the membrane replacement mechanism is that active neurons sequester a protein tag, that tells the membrane replacement mechanism to add another ion channel. The ion channel is added and the tag is removed so the synapse needs to earn it again to keep increasing its weight. Because the tag is removed by the membrane replacement Mechanism, it is called an ephemeral tag. Now we don't know how many ephemeral tags each synapse can sequester, so we don't know the leverage that the synapse can have in a single membrane replacement pass, but, while the synapse has the tag, it is identified as an active synapse by the tag. The cell uses this flag on the synapse to do chemical functions that favor active synapses.
Despite all this, Synaptic Memory that is un-enhanced has a relatively short shelf-life. Synaptic weights are dependent to some extent on the pre-synaptic rules that determine which synapses get the tags for which memories. These in turn are dependent on when the pre-synaptic bud depolarizes, and so, anything like habituation reduces the life of the synaptic weight by limiting the number of times the cell can fire in a row. If we understand this, then we can understand that other mechanisms like facilitation that increase the number of times a cell can fire in a row, also increasing the length of the memory it can store. However the longest we can expect to preserve a specific synaptic memory is about 2 months. Ideally from the point of view of an organism, the ability to store a memory should last a life-time if the memory is important enough. so synapses by themselves are not enough.
Now it makes sense that if a synapse probably only lasts for a maximum of 2 months, that you might want to grow new synapses and train them up so that there is always a good synapse waiting to grab the signal if it ever comes. This is called Synaptogenesis. And since synapses are on fibrils called either dendrites or Axons, you might want to increase the local number of fibrils that link one neuron to another. We naturally assume that the half-life of a fibril is much longer than the half life of a synapse, and so the result should be a more permanent bond between the two cells.
Eric R. Kandel, and his contemporaries at Columbia University have suggested that there is a cellular mechanism triggered by activation of certain synapses, that triggers chemical cascade reactions, that both extend the life of active signals in a cell, and start mechanisms that grow new synapses and fibrils. He describes this as a discussion between synapses and Genes, because genetic expression is required for anything as complex as synaptogenesis or fibril growth.
Let us postulate that any neuron with sufficiently advanced synapses can perform long-term memory. Let us further postulate that all Pyramidal Neurons have sufficiently advanced synapses. If this is true, then all Pyramidal Neurons are sophisticated enough to have long-term memory. This is probably why most memory systems we are discussing involve pyramidal neurons.
If we look at the pyramidal neuron we see that the reason for its shape is a dendrite at the bottom of the soma of the cell that seems to service the Axon of the cell. It should be no surprise that the synapses on that dendrite are calcium ion channels and that part of their role is to Facilitate the neuron overcoming the natural tendency for the axon to habituate. An interesting aside is that some of the most sophisticated synapses are calcium ion channel synapses. Another interesting thing is that the synapses in the basal dendrite of a pyramidal cell, produce a secondary messenger chemical called cAMP. Between the Calcium and the cAMP secondary messenger the chemical cascade reaction begins.
So while we don't know much more than this, either the S synapse or the NMDA synapse can trigger long term memory by stimulating synaptic and fibril growth. they are of course limited somewhat in that the growth requires a secondary signal called a growth factor and that over the life of the individual the amount of growth factor gradually dwindles, and so over the life of the organism the number of new fibrils formed also dwindles.
Conditioned or Learned Reflex
Instinctively we turn towards the sound of glass breaking. But that response is not pure instinct, because glass breaking is not a natural factor in the history of pre-hominid species that didn't have glass. It however falls within the rubrick of distressing sounds. So did we learn to turn at the sound of glass breaking, or did our natural penchant to turn at the sound of anything breaking, roll over into turning at the sound of glass breaking? Alternately did we learn to turn at the sound of glass breaking by association of the sound with an already existing reflex to turn at the sound of anything breaking, by association of the sound of breaking of glass with the general case of anything breaking? We don't know, but it is a general case that anybody will turn at the sound of glass breaking unless they expected it, and were prepared for it.
Much of the work of the Behaviorist Psychologists was an attempt to figure out how we learned reflexes, or how they were conditioned from other reflexes that were instinctive. Conditioning comes in a couple of different flavors, and I really could care less, but wanted to include them for purposes of completeness. They are after all a form of memory
The first 3 to 4 layers of the Cerebral Cortex and the 6th layer of the cerebral cortex, form a special type of memory called Implicit Memory. This is a Content Addressable Memory based on a Neural Network that results in output in a manner called a Quale, or Qualia in plural. Essentially it is a Voluntary Memory where the stimulus, triggers a wealth of related output, all in parallel. this is modified by the relationship between the first three layers and the sixth layer into clusters of output called neural groups, that act together to form a single output per group.
This has the effect of narrowing down the output field about 100 fold, but also means that the Cerebral cortex has more selectivity at the neuron level than it allows at the neural group layer. the main requirement is that because this memory is content addressable and phenomenal in output, it cannot be separated into distinct memories, but must be dealt with as a whole. The data field is presented I believe to the Basal Ganglia where it is analyzed to trigger instinctual actions like turning towards the breaking of glass, and separated into clusters called functional clusters that are probably based on the need to orient towards a specific stimulus that has been detected by the Basal Ganglia, not the content of any specific object in the field.
The implicit attention system allows these functional clusters to be presented one at a time to the bottleneck where they are converted into clumps. Essentially lists of Mini-Column addresses based on an interface introduced in the 5th layer of the Cerebral Cortex.
The Clumps represent the functional clusters in a retrievable manner, and so by storing the clump any functional cluster can be reproduced. However the Functional cluster is not all that interesting to the brain, because it can't be manipulated in its Qualar Form. what is interesting is that if some of the mini-column addresses in the list are suppressed, the remainder might be more salient than the original functional cluster. So if you project the quale output onto a salience meter, and try different combinations of the mini-column addresses you might find sub-elements of the data field that are more salient than the functional cluster by itself.
When you find such a salient sub-element, you pass the resulting quale through the bottleneck to produce another explicit memory clump. thereafter you can deal with the more salient data, and so deal with the data elements within the functional cluster according to their salience, despite the fact that the actual quale is indigestible. As the system gets more sophisticated you begin to develop strategies to winkle out the most salient information faster and faster by recognizing patterns in salience.
The next step once you have discovered salient sub-elements in your functional clusters, is to start to build an index of where these sub-elements were found in the cerebral cortex (according to the list of mini-column addresses in their clump.) as a result of this, whenever you recognize the salience pattern of a particular element, you can trigger the memory of that element without having to search for the clump. Now part of this trick is to know that the cerebral cortex already has the information about that salience pattern, and to substitute a known clump for the one you were attempting to find. To do that, you have to tell yourself that you know the element. Which means that you have to recognize the salience pattern, and get a response that says, yes I recognize that pattern and know what it is. We have learned recently to call this the "Feeling of Knowing" or FOK, the signal probably indicates that a quick search of the index, turned up a similar answer.
Now we have a demand memory. Simply look a memory up in the index, know that you know it, and present the memory to the cerebral cortex as a set of pre-activations at the mini-column level, and that results in the recreation of the same Quale. Or at least that is the theory, but as Dr. Edelman has pointed out to us, in The remembered present what we are actually doing is triggering explicit activity of a constantly learning network so that the output when it is recieved is not the original memory, but how that same area of memory is stored now that it has been influenced by new knowledge.
but because the memory is always changing, there is a good chance that some of those memories are going to move from one mini-column to another and may even migrate out of the neural group they were originally found in. If this happens the index is wrong, and we get the effect affectionately called TOT or Tip of the Tongue where you know you know it, but can't find it while you attempt to search for it, using related storage strategies.
Another type of memory is episodal memory which it is thought is the role of the hippocampus. Essentially this is how we find things in time and space. What we know at least with rats, is that certain locations within the hippocampus light up whenever we are located somewhere we recognize. these place cells, are mapped onto the hippocampus by a set of mapping signals produced by the Entorhinal Cortex. Another side of the Entorhinal cortex takes the results of the hippocampus processing and makes it available to the DorsoLateral PFC.
The last type of memory I will be discussing here is the memory in the Cerebellum, where it is thought skill type memory is kept. There is evidence of a connection between the Cerebellum and the bottom-up attention system, suggesting that the thalamus is the mechanism whereby certain patterns of pre-activation are replaced with other pre-activations for actions that are more effective, thus performing skills rather than raw movements.
Marr saw the Cerebellum as being a storage location that returned sequences of signals instead of the single movements that the DorsoLateral PFC would have chosen. Thus the DorsoLateral PFC doesn't have to be involved, in skilled work, only the latero-Ventral PFC.