Question
Phenomenality of Neural Network Models
Can anyone comment on Jerry Fodors contention that neural networks are phenomenal because the address of data within the network is determined by the state of the network at the time the memory was stored?
All Answers (48)
-
If you accept this philosophical point, then it makes sense that the address of any memory in a phenomenal system such as a neural network is indeterminate. This places some rather interesting constraints on the nature of the brain, if it is to deal with specific memories. -
You have a valid point there, but how does a stimulus become context related so that it can be addressed in that manner? My interpretations of Marr's theory back in 1970 was that the stimulus was taught to the pyramidal cells by repetition of the stimulus being detected by the mossy dendrite neurons, who detect it by competitive activation. No opportunity to gather context there, it is a fairly straightforward relationship, one set of cells finds the pattern and teaches another set of cells to fires when theat pattern is detected. This is the nature of the phenomenally implicit memory. Yet evidence abounds that human memory is more complex than that, how do we bridge the phenomal limitations the Jerry Fodor discovered acted as a limit to phenomenal representation? I think that the answer has to be some form of redescription to at least a less implicit form that might indeed involve association of related context Added as an afterthought to the original storage mechanisms.
-
Ah... I think I see what you are meaning.
But I think you are misinterpreting the concept of memory to assume a demand type memory. Yes, for sure the process of recalling a memory results in a new memory, but this new memory must be in a different part of the brain than the original memory. I work with a 4 layer model.
1. implicit memory: Inductive memory sourced from stimuli mainly, content addressable and
cannot be demanded directly. (that is why it is called implicit) but becomes part of any explicit
memory because of the nature of Phenomenal Explicit Memory
2. Explicit memory: Redescription of above, that can be addressed with a place-code address and
therefore can be demanded by supplying the address
3. Skill memory that learns from explicit memory how to form sequences of functions/actions
4. Declarative memory that combines Explicit Memory with a Meta-Index and Meta cognitive signals
such as the Feeling of knowing and the feeling of familiarity not to speak of the Tip of the
tongue feelnig when the index gets our of sync with phenomenal memory locations.
Eidic or Photographic memory has more to do with the nature of Neural Group formation I think than with the model of how memory is formed. If neural group formations are fairly stable, declarative memory will always reference the same place-code address for the same information. Since the stability of the neural group formations is critical to the stability of the location of the memory within the implicit memory where it is stored, the stability of the explicit address is determined by the stability of the neural groups.
-
Graeme;
Well, the 4 kinds of memory models that you work on is very similar to the model defined by Wozniak(1990), and that's one of the standard version of working with biomemory. But the addressable context or 'pure' localization of memory to some specific areas of the brain is often troublesome, i don't say that's the only model. Consider when you are saying "Yes, for sure the process of recalling a memory results in a new memory"- do indeed create a virtual meta memory which gets erased once the recall function is done.
But i think, there must be something different mechanism working behind when we try to conceive relations between information and meaning. First, our brain regions, say, cortical neurons are stimulated when we learn something new, which have both electrical and biochemical basis. That is, when we are learning, absorbing information, finding out relations between them(thinking), and then trying to figure our what we have learned, actually, those same 4 stages of cognitive functions.
Yet, previous research has shown that it is very difficult to pin point memory location in our brain, due to the diffusion of perception and stimulation of multiple associative areas of the human brain. So, that addressable component, one similar like registry memory in computer, may be difficult to understand in such context. But it may be true for implicit memory. We learn mostly in terms of conditioning, and by repeat learning, we develop short to long term memory, so, it is not possible to develop long term memory without developing short term first. That's what common theory say and we all agree. Now, due to diffuse localization of perceived patterns of information and multiple cortical area being stimulated EVEN to recall, parts of brain responsible for storage and recollection of memory seems to overlap, i.e., temporal lobe, limbic system, hippocampus, Papez circuits in the limbic area, neocortex etc. The stimulus function of neurons of inferior temporal cortex that perceive visual stimuli also may have other functions as well. So, identifying some definite specific areas in the brain for memory is a difficult job.
If you consider animals decorticated or with lesion injuries to even other non-associative parts of the brain concerned with memory functions, one may still observe severe loss of memory and retention capability dysfunctions. In patients with Korsakoff's Syndrome, an effect due to alcohol injury to neurons, memory loss is profound with anterograde amnesia as yet also, specific lesion or injury to temporal lobes and limbic system cause serious memory deficits. That's the real paradox i think. -
Ok, so let me suggest how the Isocortical columns might cause the memory to act as if it were place-code addressable without actually having stable places to code.
In my Dual Mode Cortex Model, I suggested that the top three layers (laminae) of isocortical tissues are much like the three layers of Allocortical tissues, and that this similarity means that more or less they do the same job.
The bottom three layers (laminae) on the other hand tend to have no direct connections to the normal horizontal and vertical networks, although the bottom most layer projects off into the white material that makes up the bulk of the brain.
If we assume that the output of the top three layers is a redundant, degenerate (not optimally coded) cloud of voluntary similarities, as triggered by the stimuli reaching the brain, then we can see that this cloud can be treated as a whole, rather than as discrete memories. The discrete memories are as Fodor suggested, not the actual data, but relationships between it and previous data, that triggered due to the similarities between the relationships within the new data, to those that were previously stored.
This makes the whole experience a functional cluster that is simply impossible in its natural form to process.
Ok, so we know that the brain acts heavily to filter the data, and one of the best ways of doing so, is to Partition the cloud into smaller clouds according to the sensory modality of the original stimuli. This is done mechanically by linking the data to the sensory modality that is needed to translate it.
We know that some of these connections cross modality borders because of the nature of some peoples ability to see the colors of sound, and the colors of numbers etc. But probably the higher level associations are required for the mapping of a number to a color, or a sound to a color. At the base level, we simply have modal storage with little cross over.
Now it is my contention that some of the backcountry connections, are linked to the limbic system via probably the basal ganglia pathways, where vertical processing linked to limbic evaluation offers an opportunity for some processing to be brought into the foreground, and other processing to be shoved back into the background, in essence prioritizing the processing, and in doing so, it also tags the processing priority. This makes it easier for the body to orient towards salient stimuli, and ignore other stimuli that might not be as salient. The feedback from this instinctual filtering, feeds into the Nucleus Accumbens and the Thalamus, and is fed back through Dr. LaBerges Triangular Attention circuit both to the Prefrontal cortex as a reference, and to the Cortex via thalamic influence as a Pre-Activation bias, that increases the strength of signals deemed more salient, and also tags the different attention zones with slightly different frequencies, so that the Ventro-Lateral PFC given the reference from the basal ganglia/thalamus can gate the more im... [more] -
We need actually about 8 hierarchical stages of processing to get from this first mapping to something representing experience. Each such remapping requires its own pass through the STM, first as an implicit data cloud, and later as an explicit CHUNK (the mapping of Neural Groups I implied earlier). And so we can actually watch the prefrontal cortex go through a rapid series of peaks (local maxima) in activation as it passes the data through the process, and can predict that nothing that does not trigger the 8th peak is "Experienced" so that we become aware of it and nothing that does not trigger the 9th peak is "Consciously Experienced" so we experience it as part of our self-awareness.
-
Graeme:
Sorry, it took me some time to relate and associate few things which you have mentioned as i was able to go through some of your previous postings and some literature reviews regarding the same topic. As it is highly debated on the nature of memory of whether it is an integrated process of biochemical, electrical and neural network phenomenon, one thing that still allude my conceptual notion is that, the model of how our memory is stored and analyzed(memory recycling process), rather than, how it is formed. Yes, there are ample evidences of content addressable inductive recall process of implicit memory, the question is, in what form or shape is it stored, when we compare it with artificial neural network(in ANN, it is a bad practice if such information patterns are memorized). But in human and animal brain, we develop memory and things are often memorized first and analyzed later( quite a bad practice of learning either). Particularly, the learning network much depends on the stability of the neural network, whether by combinations of symbols, images, logic or patterns.
Particularly, the information about the objects, say a photograph or even a piece of music is definitely stored as a long-term memory, even if it happens to be seen once just for a moment. But when we start to learn new things or say, read new raw data, we require attention that must be strong enough to help us memorise written information, or conceptualize such without memorizing those information that indeed, creates memory. Yet, some persons show exceptional information retention capabilities and there some who cannot memorize even their own birth dates. Those i suppose you meant declarative or skill memory, as good recalling of previous information is considered a skilled job. Even then, as you have pointed out, process of recalling memory results in a new memory, and so, there is definite role of weak attentive process. The role of attention circuits in the PFC has got some important role in learning information and raw data, or even capturing image patterns. So, there must be similar, yet analogous process for storing images and somewhat different process in storing written information, phrases or sentences. As you have mentioned, those isocortical layers or laminae has definite roles to play in processing and retention / recalling of biomemory. But there are enough reasons why such assumptions are considered important.
Even a dumb, suppose like me, can remember tunes and movie clips heard long before(even when i was a child), but often find it difficult to remember or recall what my Professors have lectured me repeatedly during my graduate school days. So, in some sense, our mind works like a technology, having different sets of thought, the best match, selection methods, filtering and projecting those mapped patterns of information in collective manner.So, how is that whole process done and what paths do they follow? By applying the concepts of neural network... [more] -
At first pass, I think I understand what you are asking. The trick, as I understand it lies in the frequency of the tagging process, being applied to the S Synapse, and thus transforming the "Activation Impulse" into a memory trigger to push the memory through into the medium, and long term cellular memory. Sessions thought that medium term memories were primarily chemical in nature, and long term memories were physical in nature, but until the chemical pathways are fully characterized we will not know whether even that characterization is accurate.
The S Synapse does a couple of things as far as I know, first it supplies a lot of calcium overcoming habitutation in a manner called facilitation, and secondly it triggers a secondary messenger, that is probably used by all the other memory pathways. This and the Protein Kinase activity probably combine to act as a medium term "Flag" of the synapses actually involved in the signal so that they can be preferentially augmented by either synapto-genesis or by process branching.
(Pruning is automatic once a new process has failed to respond with new inputs.)
Thus, merely triggering the pre-activation signal with the tag attached, tends to trigger the storage. All storage previous to this is more or less volatile storage and must not be assumed to be permanent. Does this answer your question? -
On the second pass, I suddenly realized that you were mixing the explicit and implicit learning requirements and that was part of the issue in understanding the theory.
The above points were made about the implicit storage system, which needs to store the data in order to have similarities to work against, in similarity selection, whether you or I think it is a good practice or not, by linking that storage to the pre-activation signals (specifically the 40hz signal that is often lost during anesthesization) implicit long-term memory makes use of the significance filters used in triggering which information should go into the STM, to decide which information should be advanced into the long term implicit memory.
Now we need to look at it from the Explicit Memory, point of view. I define explicit memory as any memory you can recall/rehearse. This creates a middle position between implicit memory and declarative memory, a useful resting point for my description of the difference between content addressable memory and place-code addressable memory.
Essentially no explicit memory reference is possible before the implicit data cloud from the implicit partitioning mechanism reaches the STM. The Ventro-Lateral prefrontal cortex, with the ACC acting as it's gatekeeper, can now absorb the implicit data cloud, via the horizontal network on the second memory pass.
The Bottleneck is needed because there is no direct mapping of implicit memory to explicit memory, and so the only way that the explicit mapping can be achieved is to query each neural group using a degenerate coding scheme, and collect from the responses an image of which neural groups are allowed to fire.
This then, is the input that allows us to create the CHUNK or address list, of neural groups that make up the data cloud as partitioned by the PFC.
I really would like to go on, but, unfortunately I am moving and have an appointment this morning. By the way, for a "Dummy" your not doing too bad, you are the first to question this theory this closely, and so are either not jumping to conclusions like everyone else, or are at least giving me some benefit of the doubt, I appreciate that. More Later. -
Thanks for appreciating my efforts, but i believe i am still half way done unless i get to fully understand the DMCM associated with memory storage and recall process in-and-out.
Well, on the first pass, your analysis did help me clear some of my doubts, but that raises some more questions out of your answers which might have touched some sensitive topics.
First, it is well understood that the place code addressable memory is located in the human isocortical columns, that lay the foundation of your cortical model of memory process.But it is also true that similarity in structural architecture does not always mean they are likely to perform similar functions. The stimulus triggered activation impulse that itself transforms into some sort of memory trigger is acceptable, but it still does not give enough clue of how our brain tags the processing priority(i.e., which things to recall first?). What methods can be adopted to tag types of neurons in the map that represent data types? That leaves a deeper issue of how we call back memory in response to stimuli that we have not been primed before. Is the process of calling back memory similar like a door bell? And, is priming essential for pruning? May be (not)...
Second, can it be assumed that the LTM are chemical in nature whilst the MTM are physical in nature, considering that permanent memories may be stored as protein molecules, but not as patterns? Since, ECS (electroconvulsive shock) deletes STM but old long term memories remain unaffected. The initial sustenance of STM before it is consolidated into LTM provides some logic that there is formation of MTM as intermediate virtual memory. But if it be considered that ECS deletes STM, then, i think its nature is not entirely chemical one, it must be in some physical form. But to assume such would be ridiculous, since electrical shocks do also damage cellular proteins and this raises further ambiguity.
The pattern retention capacity of human brain is remarkable, since it is possible to retrace previous formed patterns in the neural groups(this contradicts LTM as viable chemical process). But suppose, thinking is not an entirely a process of pattern matching, as it may rather be increased firing and reorganizing reverberation of neuronal impulses(if one still go by Hebbs theory). So assuming this, the chemical pathways of memory characterization is dependent on the physical process and there is a possibility of reciprocity. So, how would you consider mapping of the address codes at the neural group level as data cloud considering it as a "physical phenomenon" and then to tag specific types of neurons in the map that would likely represent such types of data? Any experimental methods possible to probe such tags, using radionucleotide fluorescent dye techniques or noninvasive Magnetoencephalography(MEG) or any other techniques?
Herein, i presuppose you are back to Fodors assertions about network addressing in brain and that makes a good point. But... [more] -
OK, somehow either I wrote it backwards or you didn't catch it, but the MTM storage was chemical and the LTM was physical. Kandel's work clearly points to process growth as an expression of long-term memory, (His assumption that the LTM was only in the hippocampus, was just a short stop on the way to this Model.
Further if you have collected much of my work, you will find that I am presenting the idea of tagging by frequency shift in the synchronous frequency bands we call brain waves. The Model suggests that each of the parallel memory loops in the brain, and perhaps other control mechanisms, have their own base frequency range associated with one of the brain wave regions, of the spectrum, The reticular formation, feeds the base frequencies via the thalamus, pons, subiculum, etc. to each parallel memory loop, as part of the partitioning algorythm for that loop. Since all the loops are centered on the cortex, the actual cortex, may show multiple frequencies simultaneously as the various parallel loops of the merely very parallel memory system vie with each other for more bandwidth.
They vie with each other at the cortex level because that is the level at which contact with the STM becomes practical. My recent work on the Nucleus Accumbens, (NAc) suggests that it acts as a load balancer between all the parallel memory loops, the control mechanisms and so on,
Although the model has recently been muddied by the contention that the Nucleus Reticularis Thalami does not indeed act as an inhibition layer between the signals of the thalamus, and the cortex, it's link to the NAc clearly indicates a steering role for this thin layer of inhibitive nerve tissue. -
On the idea of STM being "Chemical Memory" I think we need to separate Ion Channel Action Potential based memory from medium term chemical pathways that are not yet converted into and therefore are not (at least yet) the growth of processes (such as dendrites).
One of the natural confusions is the fact that all memory at some level is chemical memory whether the chemistry has to do with the synapse, MTM, or the growth of dendrites.
Part of the confusion in the LTP issue was the fact that there were at least three terms of DNA expression depending on which NMDA cells were being studied.
For instance the Early DNA, and the Late DNA bracketed a third DNA based process that was just being studied the last time I looked at this phenomena.
Further there was some redundancy in the chemical pathways thought to store the memories, and some interesting interference patterns that for instance allowed competing memories to interfere with the storage of Short Term Memories (in the cell), I should note that cellular STM is not the same as the STM found in the Ventro-Lateral PFC. Cellular STM probably has a life-time of 3 seconds before it habituates, while VLPFC STM can (but seldom does) hold memories almost indefinitely. That might be why many call the VLPFC Working Memory rather than STM.
Protein Sequestration at the synapse, in my mind, is simply a place holder that tells the learning chemical pathways which synapses are involved with the activation of the cell. My guess is that there are either two types of proteins involved (types as classes I mean not variants) or that there is some way that some of the learning mechanisms can turn off the protein pruning mechanism in the membrane replacement mechanism, giving us two classes of sequestered proteins, prunable, and survivable proteins.
If we know about the bio-chemistry of the sequestration process, we might be able to find some chemical combination that is only there in prunable or survivable protein sequestration, and thus determine whether a particular cell is using a particular learning technique. -
When you start talking about electrical shock treatments, I am of course out of my depth. However I would assume that the shock treatments affect first the ion channels, because they overpower the channels resistance, and thus change the ion balance of the cell. There will be of course some damage to the membrane of the cell as well, and perhaps it will affect the sequestration process, and change the survivability of the protein tagging. Any structural growth will be unaffected, even though the individual synapses on that growth will be affected. As a side issue, membrane replacement mechanisms will attempt to contain the damage and in doing so, may change the priority of the synapses in their attempts to change out destroyed proteins.
-
Very interesting. Well, as i was going through all those you wrote that makes up quite a huge volume and so, difficult for me to get through all at a time, the reason i might have missed some of those previous writeups. Still, i find you have already mentioned about these topics in some detail.
However, i may need to distinguish first between the layers of memory and the layers of the cerebral cortex responsible for those different kinds of memory. This will further enhance our knowledge of how biomemory is formed and where it gets stored. The architectonic model of brain parts associated with storage and recall process reminds me of a classic work - "Architectonics of the human telencephalic cortex", by Braak. In that book, the author mentioned about the origin of archicortex and delineated clearly the sensory modes represented by brain. There is also some short description of linking short term memory via horizontal network(Laminae I), as well, about the attention system and STM processing units of the brain which much depends on the functioning of primary sensory area.
So, let me do a quick recap from some of those mentioned in your articles: I first organize the basic structural areas associated with specific functional domains:
____________________________________________
Structures. | Functions
-----------------------------------------------------------------
| Frontal lobe. | Smell
| Occipital lobe. | Vision
| Brainstem. | Balance and vision
| Temporal lobe. | Sound.
|. Parietal lobe. | Proprioception
|. Vomeronasal organ. | Smell and taste
| Striatum & Basal ganglia | Reactive conditioning system
| Hippocampus, NAc. | Memory
------------------------------------------------------------------
_________Structure-Function Relationships_________
The comparative structural and functional composition of different areas of the brain having differences in their degree of condensation of activation and inhibition phases which formalizes models of memory formation, sensory perception and information processing is an important aspect to understand such models in details. Recalling your 'Weak Attention Model', that is based on memory phase activation by the thalamic process to secondary access memory that results from state change in NAc, depends on the phase sensitive cells of the cortical laminae. You have also specified some roles played by the mini columns, pyramidal cells and ventro-lateral PFC frequency and phase sensitive cells that acts as gating signals to the STM. That memories can be mapped from addressable locations spanning over specialized circuits based on which you have proposed similarity detection model is... [more] -
Ok, lets try a third time to write a response to this...
One aspect that I see is that you have confused the volmero-nasal organ with the olfactory bulb. The olfactory bulb is of course involved in both smell and taste the volmero-nasal organ only with Phermones.
In fact the Volmero-nasal organ is NOT consciously accessible as a sense, but seems to link directly to the instinctual centers instead of going through the explicit memory system.
Now one of the things that I was talking about earlier was the functions of the laminae which are not linked directly to any organ, but tend to have a structure all their own.
Consider this:
Laminae I: Connective/Training area for the parallel Horizontal memory cycle.
Laminae II: Horizontal Storage/Network Interface level
Laminae III: Vertical Storage/Network Interface level (Transmits to each enchephalon).
Laminae IV: Thalamic Data Link Layer, receives from the thalamus sensory data
Laminae V: Thalamic Pre-Activation Layer, Conditions the preactivation and synchronous frequency signals needed to implement place code addressing Partitioning and LTM actvation.
Laminae VI: Column Management, manages the signal/noise properties and creates an opportunity for Containment of a number of Mini-Columns so that a degenerate coding regime can access a more stable subset of data.
Now as far as I can see this function block remains more or less stable, the actual encephalons that are represented in each Tissue may change, as may the internal connectomics within and between homogeneous tissues, but this means that more or less all columns are doing the same job, it is at the tissue level we should expect variations. The main interest I have in the columns is determining at what data-level, the connectomics are arriving and leaving the column, since this helps with identification of what type of connection they are making.
For instance, the existence of a third laminae Giant Pyramidal cell, when filtered through my understanding of pyramidal cell size versus connection distance, indicates a connection to the mesencephalon, that seems radical in the case of some multi-modal associative tissues, but indicates to me, that skill memory is not linked directly to only action but also processing. -
Another place where connectomics comes into play, is the Belt area, where connectomics between the columns leads me to believe that the columns are active in response to their neighbors in the Core. What I interpret this to be, is a chance for the memories of the SUB-CHUNKS to be linked to memories of the past sub-chunks which in turn, gives us a bit of a historically active link that tells us I think whether a memory has been primed or not. Primed memories can then go on to the Associative Memory, but unprimed ones need more processing before they can be accepted.
The second maxima in the pfc activity train, has therefore to do with how familiar we are with the nature of the sub-chunks that are passing through our brain. Part of what we are doing during this period, is trying out different combinations of Neural Group address lists based on the original functional cluster list, to see which ones are pre-primed, and which ones aren't. A totally novel data cluster, will have few commonalities with previous memories, which makes it a problem to be solved by eventual linkage to experential processing and to possibly conscious processing. Familiar data clouds will quickly resolve into existing data clusters associated with salient sub-chunks. -
I think part of the confusion related to familiarity lies in the fact that priming is thought to be done at the implicit level, and here I am talking about an explicit function, that none-the less evaluates the implicit values. I must note here that the Explicit memory I discuss here is a second stage memory, evaluating essentially an implicit data set, while the explicit nature of the process, is ignored by classical definitions. It, you could say, is the mapping function that maps between implicit meaning and explicit meaning. Without sub-chunk analysis, there would be no explicit meaning for most of memory.
I hold this to be a definition problem related to classification, rather than to any error on my part, since the mechanism seems at least to me to be clearly indicated by the connectomics of the Belt. Admittedly I am basing my work on an early "Classic" as you noted previously.
There are no doubt later books that I could, if I could find them, work with, but it took years for me to find the terms Architectonics and Connectomics, and I only had access to the book involved for about 2 weeks, since it wasn't available locally and had to be brought in on inter-library loan.
Since at least here in the Western World, Academics delight in changing definitions and jargon, the time I might have to spend to get access to such a similar but more up-to-date book, might exceed the time I have to spend, to access it, if you know what I mean. I just watched a YouTube Video that claimed that the Basal Ganglia were not to be called ganglia anymore but to be called Basal Nuceli. There are centuries of work calling them ganglia, what possible benefit except to make the teacher more influential, or to give the students a school related way to tell their former classmates from the great unannointed, would there be for changing the definition?
OH yeah the reason given, Ganglia lie outside the brain... Where the inside of a snails brain and the outside are, must be really difficult to determine because the snail tissue was called a ganglia. -
Graeme;
Well, back again, did a bit of literature search what you meant by those functional clusters and the priming of memory, even if i consider those as episodic perceptions or repetitive tasks.What comes out interesting is the fact that the anatomical peculiarities of tract fibres crossing over the cortical areas, and the neural groups associated with memory formation have a very complex organization. Mapping of those neural groups functionally would yield direct evidences of priming and storage i suppose, and is being one of the hot topics in localization of memory in the brain. Well, there are ways of fibre tracking methods that apply probability maps to locale tracts associated with information processing and pattern recalling.
Apart from the PFC, i also find there is some roles being played by PMC (M1) and the pyramidal tracts involved in processing of afferent perceptual information while the anatomical connectivity seems to be important in those cases. For example, the cellular components of memory cluster formation is generally based on the topology of neural group distribution in the brain. Not only the core areas of the cortex, but as also, the ascending tracts through the brain stem that terminate in the thalamus also play some definite role in central cortical coding theory as of how perceptual representations are coded, decoded and their patterns stored in neural group clusters. For this, one would require assumptions for a large scale modeling of network functions and connectivities involving the central neurons(connectonomies i suppose), interneurons and synapses in the background. So, in some way, identifying the neurons or neural groups involved in rhythmic behavioral pattern generation can be studied physiologically. For, that one would need to delve deeper into the understanding of the microstructure of the neural groups in great detail.
The nature of inputs, if idealized, could definitely provide clues about the nature of the layers of information processing and outputs and some approximations are thus, required, which i previously thought could be avoided, as too much assumption of parameters which i thought could lead to noisy outputs, if one models similar small scale simulation of networks computationally. The optimal number of inputs into a neural network if known without much assumptions, could help model basic behavioral pattern maps that can be studied and compared on a time series scale. The problem however, lies with 'those sure shot "optimal number"' and gaining detailed maps from such deep seated systems like the limbic system seems much difficult tasks. However, noninvasive techniques of mapping and modeling such connections are evolving fast i suppose. So, identifying anatomical parameters and physiological characteristics of group neurons is important, if not a prerequisite. I have a point to specify regarding such specific localization of memory traces in the cortical pathways, the cortex and the brain... [more] -
I see I have been remiss in indicating the nature of the Distributed Interpreter Model, I use. In essence I use a Forth-Like dual loop interpreter with optimization as the base model.
The Inner Loop, which in the Forth Computer Language is actually the implementation of the language macros, essentially converts macro-calls to command lists, in the base language usually assembler.
I see this as being the cerebellar function, in essence pseudo-sequences consisting of activations of higher associative and motor areas, make up the base language. These are served up by the cerebellum on a contextual basis, and represent a tactical level of involvement. The outer loop in Forth follows the actual program, and converts from macro-names to macro-calls.
I see this as being the function of the SMA, and that the sequences built up in SMA functions are also stored in the cerebellum, as macros, so that if the right context is delivered, the right program or automation is triggered. What I am not sure of, is how the SMA program is selected by the prefrontal cortex if that at all happens. Except that the SMA is closely linked to the Orbital prefrontal cortex, making it a natural for Model-test-execute functions. Given an associative activation and a data activation at about the same time, it seems reasonable that the data activation will be processed via the associative groups involved with the result that an equivalent to a data processing command will have been achieved. I call this technique complicit attention to separate it from implicit attention and explicit attention. The assumption being that the data output feeds the corpus collosum, and the corpus collosum feeds the higher associative areas. -
In this design, the distributed but semi-specialized higher associative areas feed back into the STM in much the same way that the cortex does, and the output from the STM is linked via the corpus collosam to the next stage of processing.
What this allows is the chaining of strategic macros, to form a stream of automation, where the number of steps is indeterminate.
In other words, the primary difference between this and a central processor lies in the distribution of processing across three or four major areas of the brain instead of concentrating it in only one central one.
As well this architecture is closer to being natural for the brain because it applies the data processing models only when parallels exist in the neural circuits. -
Graeme;
Few things are clearly apparent from the discussion that you have presented and provides one with an idea with an intention that indeed, there must be a good theory behind a model on which to work upon, or even base one's experimental acumen with good theoretical overlay. In that, your work primarily concentrates on some new methods that you describe as IMSD, DIM , Complicit Attention (CA)and the MTE models. That gives an impression that your research is much oriented toward program automation (PA)based on contextual inputs with coding implementation if i am somewhat correct and you have included some aspects of optimization as well, as if you have mentioned in your previous analysis.
So...and yet again, "optimization"...well ahh... now it sounds really good and hence, not always a suboptimal approach. I would draw one more supposition that you have considered-the associative activation and data activation quite well, based on the theory of attention which you have developed the so called complicit attention (CA) model. Well, this new term do not exist in modern or classical literature of attention models(barring implicit and explicit) as far as i can comprehend, and now, it seems that you have an eagerness to move beyond such conventional paradigms defined in the literature(i did a thorough google search, but couldn't find in exactitude your model of complicit attention and MTE as well as yet) . So this seems more interesting.
In that sense that such an attempt to model macroscopic behavior from microscopic elemental objectivity do indeed aid oneself to get some detailed understandings of the working environments of a system, whether small or big. The question thereof arise, whether your model is just a representation of a biological system(computational) or you're more inclined toward computational(biological) aspects, or simply computational analysis based on imaginary data?
More important, there are certain limits to reductionism. But indeed, circuit rewriting and rewiring definitely require some degree of reductionism as for example, analyzing emergent behavior of circuits to inputs and their output patterns whether if satisfactory or not. It is often talked about in modern scientific literature regarding biological systems approach analysis preceding computational modeling or computational modeling of theory as an essential ingredient to model biological phenomenon. The vista of computational models, as well as softwares, are now cheaply available and in vague, but what we are perhaps missing is-- those aspects of classical experimental outcomes based on intuitional model building, testing, and the acts of serendipity. And here, i find there is some definite application of intuitional approach from your end. However, it is equally and conceivably remarkable that coupling of computational models with biological studies do yield richer facts underlying cellular dynamics, and many older assumptions and theories are now being rework... [more] -
OK, as usual, when you post you bury me....
Sidharta: "Few things are clearly apparent from the discussion that you have presented and provides one with an idea with an intention that indeed, there must be a good theory behind a model on which to work upon, or even base one's experimental acumen with good theoretical overlay. In that, your work primarily concentrates on some new methods that you describe as IMSD, DIM , Complicit Attention (CA)and the MTE models. That gives an impression that your research is much oriented toward program automation (PA)based on contextual inputs with coding implementation if i am somewhat correct and you have included some aspects of optimization as well, as if you have mentioned in your previous analysis."
Oh, boy, now I have to learn new contractions for my theoretical models...
IMSD?? Something about implicit memory??
DIM?? Is this perhaps my Dual Port Memory model? I usually call it Dual Ported Cortex model, so would that be DPC?
MTE?? You got me on that one, I don't remember ever using that contraction.
Ok, now to answer the PA question is my theoretical model a PA based model?
I think the brain has a self-programming interface to itself, certainly. What I am describing however is an ad-hoc macro creation capability linked to an optimization strategy that causes macros to converge on efficient tactical units despite their original ad-hoc, rote learning mechanism. I did not by any manner of means describe the whole system, just indicated that while it followed a Forth like inner/outer loop like structure, there was also an optimization mechanism involved, which I believe is the source of the convergence that makes for instance, language possible. One thing I am trying not to do is overspeed my knowledge base.
Sidharta:"So...and yet again, "optimization"...well ahh... now it sounds really good and hence, not always a suboptimal approach. I would draw one more supposition that you have considered-the associative activation and data activation quite well, based on the theory of attention which you have developed the so called complicit attention (CA) model. Well, this new term do not exist in modern or classical literature of attention models(barring implicit and explicit) as far as i can comprehend, and now, it seems that you have an eagerness to move beyond such conventional paradigms defined in the literature(i did a thorough google search, but couldn't find in exactitude your model of complicit attention and MTE as well as yet) . So this seems more interesting."
Ah, I think I see. Optimization is a "Black Art" to many. I have read up on compiler design, and while I am nowise an expert, I think I understand the necessities of Statistical Optimization techniques from a work-flow perspective.
Complicit Attention is perhaps a hokey name, but it describes a model of how a voluntary implicit memory system can become a directed processing system, which by no means makes it a computer.
The critical aspects... [more] -
Graeme;
Thanks for your much valued explanation. I feel sorry if i gave you much stress by burdening you with loads of questions...¤¿
no no, actually i quickly generalized your models into my own version of acronyms as often i am fighting with browswer problems, several lines vanished from my posting as soon as i often press submit-as my broswer is not a standard version, andso, that's got nothing to do or take much pain confusing your own models.
Well, the critical ideas presented by you are however, bound to generate intellectual confrontations, and so, its a pleasurable environment to be in such a formal yet academic discussion with you. And its also a good opportunity to go deeper in discussion that would bore more information regarding the topics of interest. As you are quite right to say that learning modules should be based on both the capacity to absorb and understand from failure to comprehend learned modules, particularly in children, that invariably would help children of this part of the world where at the tender age where they should be playing and learning all the fun of informational materials, are forced or bound to perform child labor practices for which they are not responsible. So, instructional material should be designed to create a greater interest in learning which should be funfilled yet embedded in serious thinking.
Well, any research of this proportion that you are carrying out all by your own effort merit definite appraisal. And indeed there is a cost attached to such endeavor, the cost of search and allied costs of materials. So, yes, finance is important and i believe it will be raining shower of funds some day i pray to God, only that you should keep the thought machine going...
I am not an expert to fully gauge the far reaching effects, but can definitely foresee some good prospects bearing down where, on the business side, some applications might be benefitted, i.e,. designing memory aiding softwares, learning aids, cognitive augmentation and such related concepts.
so, that day may not be too far you really don't know how funds may flow in. In such parlor, i understand that there have been drastic austerity measures regarding scientific fundings, as some of my fellow colleagues are getting bedlocked with limited finance. But i hope as you have come this far, you have the ability to go far beyond and your findings are taking good shape.
There are some research funding sites for basic research;
The Society for Amateur scientists
You may know Dan Guitierrez and his citizen funding activities better
and;
Fondation Les Treiles, --an organization that funds partially in cognitive science domain. You may send down a proposal to see if such funding opport. are there.
Also,
Global Philanthropy Forum. -
By way, i find it difficult to keep my curiosity to know the facts underlying scientific processes at the bay. So, if you can find answer to explain what did you meant by "by the state of network" you mentioned in the topic head question? Does it mean the physical state or the state of activation? How can one describe the state of a biological network? It can be either in stimulated state or salient state. Or any other?
-
When I am talking about a "State" of the network, I mean the conditions of the networks activation just before storage. One must understand that a network, can be said to have a state, or any number of states depending on prior activation. These states, feed back into the storage of new states, by the simple fact that they already exist, and that the implicit nature of the new state is to some extent recorded as variations on the previous state. What this means is that unless the order of the states is exactly the same, and even then because of the uncertainty involved, no two network implementations of the same stream of memories will be exactly the same.
This is one reason why we need optimization after the fact, it allows the memory to stabilize conceptually drawing to some level of convergence on say what a particular word means, for instance.
Without this convergence language would be an individual thing with no real cross-over. Dogs for instance might have less convergence than humans which is why they don't learn language as easily as humans do. -
which means something depends on something else. But you are speaking about causality which means that one functional set of states is dependent on the nature of architectonic states of the neural network. In such sense, the strength and veracity of the neural network depends on the strength of connection and activation state of the network. This i think would refer to neural chains whose connectonomics depend on signal modulation. This is different from neuroplastic activity i suppose since flow of information across such a NN is not always unidirectional. The question is, whether states determine the nature of network or the nature of information being processed define neural states that in turn affect phenomenality is perhaps difficult to comprehend. Getting little confused.
There should be generally a time lapse between decoding information processed during synaptic transmission and recoding the same across the channel and the conduction specificity across synaptic channels are determined by polarity of axonal membranes, the compositional strength of information being demanded to be processed. So, the states you have mentioned is actually what metastate- i mean the nature of the state.
There are actually some examples or researches on such temporal coding and neural states.
"Computing without stable states: A new framework for neural computation based on perturbations"
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.131.277&rep=rep1&type=pdf -
OK, let me try to respond to this, because your confusion is catchy.
Sidharta:"which means something depends on something else. But you are speaking about causality which means that one functional set of states is dependent on the nature of architectonic states of the neural network. In such sense, the strength and veracity of the neural network depends on the strength of connection and activation state of the network. This i think would refer to neural chains whose connectonomics depend on signal modulation. This is different from neuroplastic activity i suppose since flow of information across such a NN is not always unidirectional. The question is, whether states determine the nature of network or the nature of information being processed define neural states that in turn affect phenomenality is perhaps difficult to comprehend. Getting little confused."
Flow across the network is definitely NOT unidirectional, in fact one of the reasons that Dr. Edelman has stressed re-entrance as part of his theory of neural darwinism is because it covers both feedforward and feedback equally.
I think we also have to think in terms of process dilation as well, It is not enough to think of just the central or primary sensory area, we have to take into account that sensory data is going to spread out from there first to the belt, then to the association areas nearest the belt, and on to the specialized areas that deal with each association, then further to the cross associative areas that link multiple sensory modalities and then the multi-modal areas that draw together information from multiple sensory modalities and so on. This flow is controlled by gatekeeping steps in the attention system, more or less releasing the natural dilating flow, as practical.
Dr. NG's work at Harvard, suggests that a minimum of 5 stages of Machine learning is needed to achieve the accuracy of human perception, and in some ways exceed it as well. However these stages are not integrated like knowledge is within the brain, and part of the reason why, is that he is using separate streams for each sensory modality and not reintegrating them like the brain is reintegrated by the linkage between the sensory modalities created by the cross, and multi-modal associations. I think at least two layers of Machine Learning just associated with integration one at the cross modal level and one at the multi-modal level (Such as the What, and Where Pathways) would set the system NG describes up for transfer to episodal memory.
Sidharta:"There should be generally a time lapse between decoding information processed during synaptic transmission and recoding the same across the channel and the conduction specificity across synaptic channels are determined by polarity of axonal membranes, the compositional strength of information being demanded to be processed. So, the states you have mentioned is actually what metastate- i mean the nature of the state."
Right, I think you have something there. Be... [more] -
I was looking at that posting of the Liquid Model for Neural Networks, and I think that it works especially well for early implicit memory (One laminae) One of the confusing factors is that pigment staining brought out a whole lot more neurons than any staining before it with the result that the number of neurons in a column has gone up by at least a factor of 10. This suggests that it is the Mini-Column that is the essential Liquid circuit, not the Column as first thought. If that is the case, then multiple mini-columns make up a column, and recruitment is not as much of a problem as suggested in the article.
In fact some redundancy would be likely for any first order liquid column, if only because the coding for the pre-activation must be degenerate in nature, and must grow in place as needed.
One way of looking at the column is that every laminae in the column is either contributing to the liquid function, to some addressing scheme, or is a read-out. It is my contention that the read-out laminae are the first three laminae allowing the liquid to output the same values whether or not the data is referenced by context or by indirect addressing (ie. Place Code).
In this model, the interneurons are the liquid, and the read-outs tend to be the larger neurons interspersed in the top three laminae. If my size relates to distance model is correct, the larger neurons are detectable mostly because they have terminations outside the column. It is an interesting model, thanks for bringing it to my attention. -
This time you really got me back into the library to search in materials i suppose is perhaps missing on the current scenario- I would like to explain to you why this has been such now what i find is that your findings and my questions may be leading us back to 2 Major directions or domains if you belive.
First,
what i have been able to deduce from your previous analysis is getting very interesting the more deeper we penetrate however in theoretical aspects:
that,
1)The directional component of vector analysis of signal flow across the neural network,
conditions of the network activation prior storage and the time compression involved at the episodal memory formation level, which offers new thoughts on entirely physical aspects( would tell you why Markov chains may be applicable at this stage when i am talking about memory formation as a discrete phenomenon following some principles, am not sure, whether Poisson process
and then,
2)The cellular and the molecular basis of memory formation, storage and retrieval as well, on the glial control of synapses as a basis for the dynamics of synapse formation.
Now, one may consider both these two dimensions in congruence regarding physical aspects and biological basis of memory formation, both are interrelated and equally important and one aspect determines how the other might act or function. As Litchman' group have shown how nonneuronal cells influence both synapse formation and elimination while Haydon's group have started to look beyond neurons to discover how glia may be involved in information processing, as you have indicated in one previous article here, which lets me to think that there are yet bigger concepts behind memory formation and synaptic army building.
I have a notion here that quantum molecular energetics may play some part with regard to teleaffinity and telesponding activities that are yet well understood, and something beyond speculation as described by those authors. Give you a reason why:
that these research groups are talking about mysterious phenomenon of how Schwann cells sense neighboring synapses are in trouble when they are severed or injured and thus sprout branches that extend to those damaged synapses even if they are distance apart. Well of certain there are a myriad of biochemical activities involved in such processes underlying chemical mediators of neurotransmission, it is equally interesting as of how nervous system changes its response through experience. This may help understand neurostatics and glia induced cellular basis of learning and memory formation.
Another aspect is the intermediate liquid states in memory formation as well how information gets integrated within the brain when you speak of sensory modality. Then, what factors determines permanent memory storage and how information flow is controlled within the neural circuitry becomes more complicated indeed. More thought follows on this subject as i am trying to dig harder on this. -
Now by assuming that the intermediate memory is in homogenous state of liquid, the force by which i mean the real momentum of information flow across the neural groups is perhaps independent to the spatial coordinates(lipids, lipoproteins and peptides) which can be defined by equations of motion in such liquid state whose probability distribution i think could be modeled. The time between each perception on arrival of information and the time taken to differentiate known patterns from randomness when you are are saying about such temporal events which is a continuous process (perception) and whose occurrence may or may not be distributed uniformly, and if that is not a random variable(event), it must have some vector parameter, so, i mentioned about the directional component of information flow across neural groups.
One must be careful enough to note that perception may be a continuous process while learning and recalling may be a discrete process, it need not be continuous, since we do not always learn, where learning is periodic activity. -
OK, here we go again. I love this...
Sidharta:"This time you really got me back into the library to search in materials i suppose is perhaps missing on the current scenario- I would like to explain to you why this has been such now what i find is that your findings and my questions may be leading us back to 2 Major directions or domains if you belive.
First,
what i have been able to deduce from your previous analysis is getting very interesting the more deeper we penetrate however in theoretical aspects:
that,
1)The directional component of vector analysis of signal flow across the neural network,
conditions of the network activation prior storage and the time compression involved at the episodal memory formation level, which offers new thoughts on entirely physical aspects( would tell you why Markov chains may be applicable at this stage when i am talking about memory formation as a discrete phenomenon following some principles, am not sure, whether Poisson process
and then,
2)The cellular and the molecular basis of memory formation, storage and retrieval as well, on the glial control of synapses as a basis for the dynamics of synapse formation.
Now, one may consider both these two dimensions in congruence regarding physical aspects and biological basis of memory formation, both are interrelated and equally important and one aspect determines how the other might act or function. As Litchman' group have shown how nonneuronal cells influence both synapse formation and elimination while Haydon's group have started to look beyond neurons to discover how glia may be involved in information processing, as you have indicated in one previous article here, which lets me to think that there are yet bigger concepts behind memory formation and synaptic army building."
OK, but as usual you are presenting them as if the later one was first. So lets start with 2, the Cellular and molecular basis of memory formation storage and retrieval. Glial assistance with this is done at the cellular level, and the start of the process at least happens the first time the STM rehearses the memory.
Over about the next 480 milliseconds the processing gets to the point where the hippocampus forms episodes, and thus we have a distinct latency before the likely processing that involves awareness, or alternately the later step that involves consciousness. It is quite practical to think that the vector that perceives is continuous, and the hippocampal evaluation is less contiguous, and may be split up into episodes, merely because the temporal compression involved means that 10 episodes can be recorded in the hippocampus while 1 is being recorded at the sensory level. This also vies for a shorter term for the hippocampal loading not forcing it to be 60-80 ms long. 8-10ms being more likely. Sorry going to have to take some time away, for errands Back later. -
Sidharta:"I have a notion here that quantum molecular energetics may play some part with regard to teleaffinity and telesponding activities that are yet well understood, and something beyond speculation as described by those authors. Give you a reason why:
that these research groups are talking about mysterious phenomenon of how Schwann cells sense neighboring synapses are in trouble when they are severed or injured and thus sprout branches that extend to those damaged synapses even if they are distance apart. Well of certain there are a myriad of biochemical activities involved in such processes underlying chemical mediators of neurotransmission, it is equally interesting as of how nervous system changes its response through experience. This may help understand neurostatics and glia induced cellular basis of learning and memory formation.
Another aspect is the intermediate liquid states in memory formation as well how information gets integrated within the brain when you speak of sensory modality. Then, what factors determines permanent memory storage and how information flow is controlled within the neural circuitry becomes more complicated indeed. More thought follows on this subject as i am trying to dig harder on this."
I can't respond about teleaffinity and telesponding activities as I am not aware of them, but, I have always wondered whether the system was sensitive enough to capture quantum events that are significant enough. Suffice it to say that I don't buy Penroses arguments about quantum level entanglements at the buckytube level. There is too much noise to signal ratio at that level to get consistent results. One or two people might be able to get better results if their readouts are better tuned to quantum level impulses, but that hardly proves anything in the general case. As to intermediate liquid states, I believe each laminae has its own liquidity, and each column is more or less insulated by inhibitive surrounds that it is only when there is damage to the inhibiting cells, that there is usually flash over.
Thus the column may share a sampling of liquid states, and their read-outs. Different sub-networks can have different timing factors, and still retain the same base liquid states simply because they have their own read-outs related to that subnet within the laminae. Readouts do not need to use the main network, and often are situated so as to readily feed data from one readout, to another liquid state in such a way as to act as a comparator, or to link processing to a more remote but not far from adjacent processing center. -
Sidharta:"Now by assuming that the intermediate memory is in homogenous state of liquid, the force by which i mean the real momentum of information flow across the neural groups is perhaps independent to the spatial coordinates(lipids, lipoproteins and peptides) which can be defined by equations of motion in such liquid state whose probability distribution i think could be modeled. The time between each perception on arrival of information and the time taken to differentiate known patterns from randomness when you are are saying about such temporal events which is a continuous process (perception) and whose occurrence may or may not be distributed uniformly, and if that is not a random variable(event), it must have some vector parameter, so, i mentioned about the directional component of information flow across neural groups.
One must be careful enough to note that perception may be a continuous process while learning and recalling may be a discrete process, it need not be continuous, since we do not always learn, where learning is periodic activity."
One thing that makes me militate against lockstep processing is the idea that different pathways or processes have slightly different physical lengths, and that part of the reason the liquid model works so good, is that these tend to create structural sensitivities to frequencies of one sort or another. The net effect, is that the timing of the network as a whole can be measured, but any attempt to define it as a lockstep state machine is going to fail on the fact that the steps cannot be but out of step, because there is no fast enough clock signal that every part gets to link them to. As such we are looking at relaxed timing where the timing per step is only approximate. It is a step in the right direction then that we can predict the completion of a particular process in the 60-80ms range, but can rewind if completion has not been achieved.
This is somewhat analogous to the Critical Block in Software, where a Critical (Usually Atomic) block of code needs to complete before the next one can be called. Using statistical optimization techniques, the critical block can be optimized in error and still perhaps save time even though it has to be rewound and a lesser technique used. -
One of the reasons I think consciousness might exist, is the necessity to rewind failed processes especially with the higher order processes of my complicit attention model. To manage this needs either a stack, or a logging mechanism that logs the macro-call needed to trigger the critical block. Because natural stacks are not part of the characteristics of neural networks, the logging mechanism is more likely, and it is likely that this is similar and reflects the episodal record but with more parameterization. When we include meta-cognitive signals as part of the parameterization we end up with feedback from the "Self" attribution mechanism being linked to the episodes where choices have been made. The juxtaposition hints that there is some link between the "Self" signal and the decisions being made, and so we find ourselves having to deal with the concept of "Will" where it is assumed that a causal relationship links the detectability of self to the decisions being made. ipso facto, the assumption is that our self made the decisions, when the self is just a signal to link attribution to the organism and has no ability to make decisions outside that attribution or cause anything but its own signal.
Of more moment to me, is the assumption that the macro-call needs to be translated by the second or outer loop, and that the critical Block can be implemented by more than one optimality. I am not sure how multiple threads of the pseudo-sequence are separated, except in a manner similar to that of an implicit memory, where the pre-activation activates the whole cluster, and the selection mechanism decides which one completes, but somehow they are either being processed in parallel, or there must be a rewind that somehow can select a previous version that worked over the current version that might not have since it failed to reach completion in time. -
As usual, when you write, there is something interesting...
"It is quite practical to think that the vector that perceives is continuous, and the hippocampal evaluation is less contiguous, and may be split up into episodes, merely because the temporal compression involved means that 10 episodes can be recorded in the hippocampus while 1 is being recorded at the sensory level. "-Graeme
S.C.: Which means that our perceptual network say, our elements of perception work in layered hierarchical model, by which i believe that the function of perception is supperposed on learning, yet both are dual aspects of information processing, one being a continuous process and the other( hippocampal) is discrete. And learning is attention specific activity as our attention network gets activated.
As we approach further toward understanding of STM and LTM at both chemical and physical pathway level, and then when we speak about neurocellular darwinism which acts to use those original states and metastates as a gateway to reinforce active and salient synaptic states, we will require somehow to know about the original states of those synapses where fuzzification occurs within those neural groups. It has been assumed by research groups that the general activation of the glial cells act to increase the strength of LTM, so glial cells being one of such factors supporting neural group firing that can be computationally simulated. Now, the discrete nature of interneuroglial conversations across the neural groups raise two important assumptions;
1)how discrete astrocyte circuits in the brain coordinate activity with neuronal circuits?
2)what is the cellular analogue of memory and bow it is affected by synaptic astrocytes? This i think you are relating to your model of memory.
Since this could shed light on formation of associations between stimuli old and new, processed by different neural groups. Suppose that if neurons function like simple telephones, astrocytes function as cellular phones, then how neural receptors are tuned to receive such broadcast to which they are not exposed before? If one would find out what kind of mediators play such role is perhaps an interesting endeavor.
When you are speaking about the cognitive process of decision making, two parallel processes are supposed to be in action. First one is the cognitive process involved in information processing, and thereof then followed by making decision based on such information. In fact we have known that our brain, in essence is a massive parallel processor. It has the capability to absorb, retain and retrieve information, we call memory. Similar things are however possible and in vague in computers. Yet, the real puzzle is that, our attention system is only capable of dealing with a single sensory-motor task, which means, we cannot do two things at the same time, unlike computers. What i may call that system in computer as paramorphic processing. The association level in brain although capable of... [more] -
S.C.:" Which means that our perceptual network say, our elements of perception work in layered hierarchical model, by which i believe that the function of perception is supperposed on learning, yet both are dual aspects of information processing, one being a continuous process and the other( hippocampal) is discrete. And learning is attention specific activity as our attention network gets activated."
I agree that the thalamic gating function makes it look like the Perception system is super-posed on the learning, but historically Perception came first so in fact the super-positional relationship is inverted. Declarative (Hippocampal) memory must be superposed on Perception rather than the other way around.
I have never quibbled about the layered hierarchical model of perception, that Jeff Foxworthy (Palm Inventor) and his group at Numanta have suggested, just the assumption that it must be tightly linked to time. As I suggested, the variation in the length of the neural processes precludes tight linkage to time.
The fact that Harvard Professor NG and his classes have shown a 5 layer hierarchical model of Machine Learning is almost equivalent to human perception leads me to think that in fact a limited hierarchical model is critical for perception.
S.C.:"As we approach further toward understanding of STM and LTM at both chemical and physical pathway level, and then when we speak about neurocellular darwinism which acts to use those original states and metastates as a gateway to reinforce active and salient synaptic states, we will require somehow to know about the original states of those synapses where fuzzification occurs within those neural groups. It has been assumed by research groups that the general activation of the glial cells act to increase the strength of LTM, so glial cells being one of such factors supporting neural group firing that can be computationally simulated. Now, the discrete nature of interneuroglial conversations across the neural groups raise two important assumptions;
1)how discrete astrocyte circuits in the brain coordinate activity with neuronal circuits?
2)what is the cellular analogue of memory and bow it is affected by synaptic astrocytes? This i think you are relating to your model of memory.
Since this could shed light on formation of associations between stimuli old and new, processed by different neural groups. Suppose that if neurons function like simple telephones, astrocytes function as cellular phones, then how neural receptors are tuned to receive such broadcast to which they are not exposed before? If one would find out what kind of mediators play such role is perhaps an interesting endeavor."
Perhaps I am wrong, but it seems to me that the sequestration of proteins in the post-synaptic patch, as part of the PK pathway, as suggested by Dr. Alkon in his book "Memories Voice" links the firing of the cell, to the active synapses at about the time it fired (Slight delay due to retrograde transmission o... [more] -
The temporary sequestered type might have a lifetime around 3 seconds while survivable types could out survive the individual components of the actual synapse because they were replaced before they could degrade due to digestion.
S.C.:"When you are speaking about the cognitive process of decision making, two parallel processes are supposed to be in action. First one is the cognitive process involved in information processing, and thereof then followed by making decision based on such information. In fact we have known that our brain, in essence is a massive parallel processor. It has the capability to absorb, retain and retrieve information, we call memory. Similar things are however possible and in vague in computers. Yet, the real puzzle is that, our attention system is only capable of dealing with a single sensory-motor task, which means, we cannot do two things at the same time, unlike computers. What i may call that system in computer as paramorphic processing. The association level in brain although capable of processing information in parallel, there are indeed limitations what i may call saliency in the context of performing multiple motor tasks. However, our brain still excels in constructing multiple representations which computers fail to do so."
Well, without going into Dennets multiple-drafts theory, we must be very careful which level of processing we are attributing multiple representations to.
Think of the brain as being a multiple layer processing system consisting at the bottom of a massively parallel processor, in the middle as a Merely Very Parallel system, and at the top in a pseudo-serial system. At the bottom we have literally many multiple representations of the same data, as it moves through the heirarchical perception system, we have an arbitrary number of different representations as we move through the merely very parallel layer, and we have at least two different reflexive images as we move through the pseudo-sequential layer. That we only perceive one image in the quasi-sequential experience at the top of this architecture simply means that it must be a simplification of the images making up the underlying architectures. -
S.C.:"Hence, the computational theories of the mind revolves around such notion of human capacity to reason, compare reception of sensory stimulation and production of motor response. So our mind is conscious and there you refer to the neural correlates of consciousness. This may however present the Cartesian ontology of dualism of the mind and embodied cognition embedding consciousness as subjective mental experience by which we all are capable to memorize events and associate events with past memory. Many a scholars have thus raised a question of whether our mind is really an epiphenomenon of materials. By what we may describe our intellect, and the phenomenon by which our brain inspects entities(Rorty, 1980) still remains a mystery. How representations mirror in our mind by reception through retina and how visual information is transcribed into memorable experiences are always interesting aspects for inquisitive analysis."
First of all, I believe that cartesian concepts of consciousness were mistaken because they assumed a "Theatre of the Mind" which we could then discover somehow. Such is not the case at all, what we might find, is a reflexiveness that lets the brain as a whole act as its own theatre, but we will not find any stage more theatrical than the STM and that because of the limitations imposed because of the bottleneck will never be able to hold enough detail to allow the depth of significance we experience.
We have to be very careful when we start talking about the Mind's ability to reason, since that plays into the pseudo-logical mindset the Greek thinkers imposed on us by inventing syllogisms. There are those who believe that the brain is actually logical, and that implies that it existed only after logic was invented, something that is patently ridiculous. However these people will cheerfully argue that anything less than a human brain (which is supposedly logical) doesn't matter and can hardly be called a brain. (A concept I heartily disagree with).
There is no question in my mind, that mind is a superphenomena of the material brain, one might call it epiphenomenal if you want, and that it is in fact an image projected back into the brain to simplify some brain circuits needed for control. That it is a simplified image, and is projected back into the brain, means that there will never be a 1:1 relationship between brain and image. -
Graeme Smith;
"I agree that the thalamic gating function makes it look like the Perception system is super-posed on the learning, but historically Perception came first so in fact the super-positional relationship is inverted. Declarative (Hippocampal) memory must be superposed on Perception rather than the other way around."-Graeme
Well i suppose i have got either myself on the wrong foot, or your explanation is pointing to an age old debate starting from Alhazen, Descartes, Hemlholtz, Berkeley and others of whether the origin of perception precedes the origin of memory formation. Anatomically, hippocampus is considered as an old world structure, while cortex particularly, the association cortex-the neocortex are considered new world structure. Architectonically, mid brain, brain stem and cerebellum evolved before higher cortical centers, if at least in primates and mammals, and also in homosapiens. Yet indeed, they inherited the perceptual functions of tactile discrimination well before it gets recorded as memory. But there are serious contentions of this "inverse problem" as it is difficult to accept straightforward that perception succeeds memory.
"We can only perceive the world around us through memory of names, understanding, and recognition. We can only remember things by first perceiving them.".....Maslow
------
Sofar, the idea seems to be complex enough to consider this debate and clear some doubts. The paradox of superposition, as you have eloquently touched the issue concerning the origin of memory and perception, seems to be a puzzling one, with no clear outcomes as yet and confounded by equally contentious debates regarding 'which came first-perception or memory?'
A variation of the clause which i think may be able to shed some genuine light on the matter if we first be able to prove that whether episodic or declarative memory originated first, since, both in most part, and more so in the case of episodic memory formation, takes place in the hippocampus, while declarative memory is now thought to be formed in the perihippocampal areas.
We perceive things first by attaching some attributes, say, some cue whether visual or auditory, to the originating stimuli. We remember what we see and feel, and yet as a child, we perceive first, then we store, and recall those stimuli which is backed up as data. So, in order to navigate the world, our brain uses our 'memorable' database to understand or match new incoming stimulus, store or discard what we like or feel memorable, using some pattern matching algorithms, and thus our natural biomemory database evolve. Memory facilitates visual perception, and by visual perception we build memory...
Well, really puzzling isn't it? Consider the simplest of the organisms which do not have any 3 of the common 5 senses that we have. Yet, they do have tactile discrimination and sense of understanding their direct real environment. By tactile perception, or by chemotaxis, they perceive attri... [more] -
Sidharta:"Well i suppose i have got either myself on the wrong foot, or your explanation is pointing to an age old debate starting from Alhazen, Descartes, Hemlholtz, Berkeley and others of whether the origin of perception precedes the origin of memory formation. Anatomically, hippocampus is considered as an old world structure, while cortex particularly, the association cortex-the neocortex are considered new world structure. Architectonically, mid brain, brain stem and cerebellum evolved before higher cortical centers, if at least in primates and mammals, and also in homosapiens. Yet indeed, they inherited the perceptual functions of tactile discrimination well before it gets recorded as memory. But there are serious contentions of this "inverse problem" as it is difficult to accept straightforward that perception succeeds memory."
Let us for the moment separate historical succession from signal progression.
It is hard to make any claims about historical succession, if only because after an organ becomes redundant it is often re-adapted to a new role, and so the old becomes new again. There is no question in my mind that the older parts of the brain had variations on implicit memory long before the cortex evolved. However these implicit memory modules were found to be nuclei of one type or another instead of flat plates such as the cortex would be if it were uncrumpled. The history of the evolution from nucleus to plate can be seen in the hippocampus where the nucleus was cleft by the formation of a different type of cell the tissue slowly giving up its spherical orientation and forming the famous seahorse shape.
Once the thalamus however became the staging point for much of the sensory information the senses more or less didn't get directed through these older routes, and some of them became redundant. New pathways grew connecting the thalamus to the now plate-like cortex, and along with them came the Isocortical tissues. Now the neural routing had to pass through the cortex before it could reach the older tissues in the hippocampus.
Now, of course there is no question of flow, Perception comes first, and declarative, and episodal memories are built on top of implicit memory structures and so must come second. Thus historical growth is reversed with respect to modern (Mammalian at least) signal flow. It is interesting that the abandonment of the brain stem related signal path for senses, caused the redundancy of the earliest brain structures, just in time for them to become much higher order structures, expanding the capacity of the memory in such strange new ways. -
Siharta:"it is difficult to accept straightforward that perception succeeds memory."
But... Perception IS memory.
One of the problems with classic management of roles within the brain is the assumption that you can separate memory from processing. While it is true that some cells are obviously more related to processing than memory, and others otherwise oriented more to transport, there is never a case where a neuron does not do all three functions at the same time. The amount of processing done in a transport neuron, and indeed the complexity of the memory that that neuron shares are necessarily reduced, but not eliminated.
This is why it is a false dichotomy to suppose that perception comes before memory, perception is all about the interplay between memory and new stimuli. and in fact in order to store anything in the way of perception, further processes need to be fed back into the memory circuits. It is not that there wasn't great discussion on these points, but simply that the discussion becomes mute if you do not dichotomize two functions of the same cell, because "Logically" those functions are separate. The brain does not make the same distinctions that Logic seems to require, possibly because the brain is what logic is meant to reflect, not what reflects logic. -
Sidharta:"=My Thoughts=
However, neuroanatomical models of memory formation and storage in the hippocampus is an old world phenomenon, considering that most of the specialized regions of the brain, particularly the neocortex responsible for higher level associative functions having 6 layers of cellular structure compared to the archicortex, hippocampus, cerebellar cortex that all have 3 layered structure(Molecular Layer I, Purkinje Cell Layer II and Granular layer III). Now, when seen from a top down approach, the neocortical cell layers late in evolution, the Layers I, II and III are specialized for sensory input processing, receive emotion-based sensory incoming stimuli (sensory cortex), whilst, the Layers V and VI gives rise to descending motor fibres which are highly developed in the motor cotex. The peculiarities and uniqueness of the layers of the cortex are thus specialized for unique functions, whether mental, or motor response. The thalamocortical layers- the primary receptive layer( somato sensory, visual and auditory) is however an old world structure corresponding to "Receptive" functions. Even if one agree that in higher forms of life, the more specialzed layers I, II, III evolved later than the deep seated mid brain structure-limbic and hypothalamus-hippocampus, the same thalamocortical layers were responsible for sensory reception in lower vertebrates. Whereas, the hippocampus proper( Ammon's horn), is an old world architecture, having much to do with both episodic and declarative memory.
The thalamocortical layers as the primary receptive layer for somatosensory, visual and auditory impulse receiver are responsible for sensory input processing and communicate with the association layers where their neurons exchange information with each other. Now, the layers I, II, and III have evolved for sensory input processing, having different histology for receiving incoming emotional stimuli. Consider that interneurons which are more abundant in sensory and association cortex that perform higher sensory integration functions, i.e., assignment of meanings. The interneurons are specifically part of neural network feed-forward -feedback controller, so, involved in central pattern generation, matching and switching functions. Anatomically, the layers V, VI cells gives rise to descending motor functions, indicating that primitive organisms and vertebrates were more concerned with lower level information processing, better motor functions and tactile discriminations. The layers of V and VI are responsible for the origin of corticospinal, corticothalamic and intracortical tracts which are highly developed in motor cortex. The layer one for reception and finer associative functions, having molecular form devoid of large fibres.
Also note the structural homology and the histological differentiation between the layers of PSC, AC and the PMC. In the layer V of primary motor cortex one may be able to find abundant Betz cells(pyramidal) in the gangl... [more] -
Once the idea of the parahippocampus being the declarative memory, is considered, it makes sense if the WHAT information gets peeled out of the WHAT, WHEN, WHERE stream at the parahippocampus and When/Where get represented at the CA3 part of the hippocampus, as long as the subiculus grabs references from both at the same time to create the What/When/Where functional cluster. when doing episodal memory, and mostly the What when doing declarative memory. As long as the Where/When entries link back to the What in declarative memory the system should work just as good.
-
Graeme;
It is reasonable to understand what you mean by the dual problem of classifying and storage of memory. That is, what memory is stored where, considering hippocampus has place codes for such. Well, to categorize memory based on the nature of perception and the process of formation could give one better idea of which memory is stored where.
You explanation generally points to assume perception as a "process" and memory as "product", to say in simple terms. The product is as much depended on the process when you have said "perception IS memory". Well, it is difficult to refute such a claim but considering perception as a process, i may ask that do all processes gives rise to memory, which is a specific product, or also do they produce other intermediates as products of information processing?
On the other hand, your analysis shed somethoughts on the "plate theory of neocortical evolution", a concept under rigorous research amongst neuroscientists which struck my mind that indeed, we are searching for a unifying theory of brain evolution that would likely incorporate both the aspects of structural evolution as well functional adaptation of the brain, in as much as to answer questions related to our above analysis, that is, on the origin of perception and memory, or say, the evolution of storage mechanisms of the mammalian brain.
It does help to look further into such aspects as cortical mapping, dynamics of cortical evolution in human and primates and to seek for answers to questions like why our brain has become so accomodatively large compared to other species, why they are structurally and functionally as well fundamentally different, and what differentiates us from primates in terms of higher cognitive processes, and lastly, why are we able to think rationally?
# that, the plate theory of neocortical evolution, which is strikingly new to me and interesting as well
#evolution of the hippocampal region and the attributions of memory
#seearch for n unifying theory to define mammalian evolution of brain, perception and memory.
Thanks for your enlightened analysis that opened up a new chapter for me, so interesting, as well puzzling, but worth considering that projected my mind on entirely new thoughts and to look intently deeper into such interesting and fascinating theories, fascinating they are indeed...!!! -
Sidharta:"Graeme;
It is reasonable to understand what you mean by the dual problem of classifying and storage of memory. That is, what memory is stored where, considering hippocampus has place codes for such. Well, to categorize memory based on the nature of perception and the process of formation could give one better idea of which memory is stored where."
I have been using the physical location, and thus the memory loop to describe which memories have been stored where. It seems to me that the 4 main memory Loops are more likely to have similar roles within them, than each to each other. I call the parallel system that is created the "Merely very Parallel" level of brain organization, because while the processing centers are still more or less massively parallel, the memory loops are merely very Parallel with respect to each other, the idea of the Nucleus Accumbens as the load balancing mechanism for the various control and memory elements brings together a ream of relatively small parallel systems into a working whole.
I think however that I am not communicating my understanding of the design correctly, if we are talking about the parahippocampus as the seat of declarative memory, and CA3 as the seat of episodal memory, then it makes sense that the brain would use the subiculus to link storage in BOTH locations, so that access of the declarative memory would index the episodal memories associated with the categorization, and accessing the episodal memories would index the declarative memories associated with the episode. One of the reasons that these circuits are so closely linked, is that they both deal with the WHAT/WHEN/WHERE data coming out of the parietal and temporal lobes. The nature of the place codes and now temporal codes found in CA3 suggest that episodal memory is more about where and when, than what. The parahippocampus might therefore fill out the pattern by being more about what than where and when.
If they are both accessed via the same pre-activation mechanism operating in the entorhinal cortex and subiculus the output would tend to include the full What/When/Where cluster of information, even though the locations of the data were in different organs. Anyways I am beginning to get a whole different feel about the nature of the processing done in the hippocampus, and it's relationship to how we "Experience" our world. I think as we understand it better we may find it the key to the correlates of consciousness.
Siharta:"You explanation generally points to assume perception as a "process" and memory as "product", to say in simple terms. The product is as much depended on the process when you have said "perception IS memory". Well, it is difficult to refute such a claim but considering perception as a process, i may ask that do all processes gives rise to memory, which is a specific product, or also do they produce other intermediates as products of information processing?"
If as I have said, all neurons have at least vestig... [more]
