Brain Physiology, Cognition and Consciousness group: topic
This is a public discussion board
Weak Attention: A 10 phase model of attention
Graeme Smith
Thursday, 06 August 2009 16:35 UTC
Attention, what is it? Some such as Hans, would have you think it is the spotlight that shines on a topic when we concentrate on it consciously. I call this Strong Attention, Others like me think that it is more a directive step in the transfer between different areas of the brain. I base my weight fairly heavily on the work of David LaBerge, who published a number of papers on something called the Triangular Circuit of Attention and Awareness. I just take his work a little further, and define somewhere around 10 phases of attention, of which 5 or 6 involve the mechanism Dr. Laberge calls the triangular circuit.
I call this less concentrated form of Attention Weak Attention, to indicate that it is not strongly linked to conscious awareness.
The main difference, and I tend to argue this even with Dr. LaBerge, lies in the assumption that attention exists only for awareness. This is a timely topic, in that current articles in Frontiers of Neuroscience discuss the MMNn signal as detected by the MEG device, and the attempts of scientists to show a link between the peak of this signal and Awareness.
My theory would have you look not only at the peak, but also at the local maximas in the MMNn signal for evidence of Attention activity, and would have you look at the last two such maximas for evidence of whether or not Awareness is created. The interesting thing is that there probably is a connection between the peak activation of the attention system and whether the last two local maximas are experienced, so the experiment would be inconclusive as to which interpretation to use.
If my model of attention is correct we can actually trace the transfer of a stimulus from memory stage to memory stage by monitoring the local maximas in the MMNn and thus monitoring the attention phases that the stimulus passes through. The model I use is a 10 phase model as I have said, even though we seldom see evidence for more than 9 local maximas. In essence each maxima denotes the transfer of data between two different locations in memory, the amount of attention activity noted for a specific stimulus is therefore the amount of transfer complexity involved in that phase of Attention.
It should be noted that The Epochs before and after each such transfer, define the periods of time that a stimulus is retained in the specific memory area involved. The fact that the last two maximas run closer together might indicate that the 9th epoch is much shorter than the others. The general signal strength probably indicates the habituation of the signal, which is why when a memory stays in one epoch longer than usual, it also tends to degrade more, and be less likely to reach conscious awareness.
This suggests that it is the process of attention that supports the signal against habituation and thus decay. If Attention is not paid to the signal approximately 12 times a second, it begins to decay before it can reach awareness. Libetts work suggests that the signal has to be protected from premature habituation for at least 500 milliseconds or it won’t reach awareness. The MMNn signal lasts for about 540 milliseconds.
In other words the lions share of attention happens before awareness is possible, supporting the Weak Attention hypothesis.
-
Replies
-
Sorry I had to break in the middle of explaining weak attention, I thought I had an important meeting, but it turned out I had the wrong week. Anyway, If there were 10 phases to attention what would they be?
The following list might explain:
- Orienting: Turning towards a salient stimulus, Probably instinctively
- Implicit Attention, Binding Cross Modal stimuli into Functional Clusters and forming Salience zones
- Functional Cluster Chunking associated with the CORE areas of the Cortex
- Isolation of Memory Elements, (Sub-Chunk) associated with the Belt areas of the Cortex
- Association of Memory Elements, associated with the Association areas of the Cortex
- Complicit Attention, associated with the specialized (Recruited) areas of the cortex
- Declarative Attention, associated with the hippocampal loop
- Skill Attention, associated with the Cerebellar loop
- Intention, associated with the SMA and the awareness buffer
- Volition, associated with the consciousness buffer
It is my contention that the last two tend to blur together in our psychology experiments, so that they are seldom separated. Thus the last two epochs are an indication of whether or not a stimuli has reached our awareness.
-
I was just over on ConerLab’s site, and was watching a video with a number of consciousness researchers in a forum discussion. Of the three I was most impressed with Patricia Churchland’s approach to Consciousness. I was more impressed with her work, than I was with Christoff Koch’s.
I think that understanding attention is critical to understanding consciousness.
My Weak Attention Model with 10 different phases of attention, could answer a number of “Hard Questions” that have accumulated in consciousness science. Characterizing Attention as being the way that data flowing from one section of the brain to another is regulated, is also helpful, although I know that it is a much different interpretation of attention than is the norm.In this theory the critical part of the attention system is the ACC which regulates the selection of information. At each step in percpetion, the ACC makes a selection according to tags that separate functional clusters of data and regulates which information from the sea of parallel outputs from the cerebral cortex is advanced on to the next stage of processing. To do this, it must have not only bottom-up preactivation, but synchronization to form the GSO tagging algorythm, and some evaluation of the data for significance, which I have learned to call salience. The tagged zones or functional clusters are then processed through a stable series of phases, as the nature of the data is evaluated.
First comes Orientation, where the body moves to present its best data gathering organs towards salient zones of the environment. Then comes the implicit stage where sensory inputs are combined with long-term implicit memory in a manner similar to content addressable memory, which causes a broad volunteer population of neurons to fire. Due to limitations of the neural structure that implements this phase, the best we can so at regulating the flow of data, at this point, is to label data that is in the same salient zone, as part of the same functional cluster.
This approach has the added effect of linking cross-modal sensory data and may be part of the binding that others see as problem. Another part of the binding is the parallelism in the output. Quite literally millions of neurons with differing data dependencies are firing as one. One of the problems with this is that there are so many neurons firing at one that the brain can’t pick one out to study at this point.
The next phase therefore is important, the data has to be converted from this wild parallelism into a more controlled parallelism that allows us some control over which elements are being dealt with. To do this we need to change the addressing mechanism by which memory elements are selected, to one that is Location-Centric, rather then content-centric. The Functional Cluster has to be addressed as a chunk, rather than as a parallel field of neurons firing all at the same time. This is where my Dual Mode Cortex hypothesis comes in.
Now all this parallelism and changing of addressing mechanisms has happened in the primary perception areas. These are known in perception as the CORE areas in each sensory modality. Data flows from these core areas into the belt areas according to the attention phase that picks out Sub-Chunk, thus the belt areas separate the data into more focussed functional clusters related to some isolatable field of the original chunk created when the data was converted from a functional cluster to a plac-code addressable form, I have chosen for historical reasons to call a chunk. The next step is to find out a bit about the data elements found in the previous step, and to do that we route the data into the Associative Areas where each data element is associated with other data elements that are similar according to some rule.
At this point the pre-stabilized data processing is over. The next stage is more like what attention is thought to be, where we start to move data around by recruitment of specialized modules. To achieve this we need a buffer like effect where we can load a data element into the brains network and feed it forward to a selected specialized processing center. While we are not sure, it is thought that the buffer regulates the passing of information via the corpus collossum. Interesting enough, we know that connections to the corpus collossum come in at the input side of most cortex tissue allowing processing of the output of the 5th phase to loop back into the first phase, as well as recruit new data specialized areas.
Now this is a point that Patricia Made in her lecture, the actual amount of the data in the brain exceeds the sensory input, and we know we are processing data that does not come as primary percpetion, when we for instance infer a line in the field of a figure even though the line was not drawn. She showed the well known figure where there are three circular pies with triangles cut out of them arranged in an triangle so that the cut outs are proportionately spaced to give the illusion of forming a line between them.
The important thing to realize is that the inferred line, is presented as primary perception even though it does not exist in the primary data. Igor Aleksander has described this as “imaginal” which is hopefully distinct in our minds from imaginary, the line is imagineary, but the reflected data that is mistaken for primary perception because it flows through the same circuits is imaginal.
Now the interesting thing about this phase where we direct data from one location to another within the brain, is the fact that to designate the data we need a chain of events that essentially describes two separate attention processes linked to form one. I call this Complicit attention for a number of reasons not the least of which is that I get to have a play on words based on the meaning of complicit. The net effect of this processing step, is that complicit attention allows an equivalent to a Command in a very basic machine language. What is important here is the dynamics of commands, not what type of processing that command does, because what it allows the brain to do is to have a process for stringing these commands together into programs, the basis of Automaticity and Automation. Up until this point, there was only regulation of data flow, now the same mechanisms provide an opportunity to automate, and run pre-programed routines. The main difference is in the architecture of the connectonics between phase 5 and phase 6.
Now we have a problem, we can move data around, and select specific elements to move, and select specific functions to process them, but we nave no way of finding where a specific element of data is. What we need is a map, an index, a way of accessing the information found in the previous steps of processing. Strangely enough we have a candidate for that, but it is encumbered by our misapprehensions about the nature of memory, so that it has two completely different roles that no one has yet reconciled.
Now let me go out on a limb and say you can’t find something unless you have a map to it, and in the brain, while there are lots of maps most of them are not associated with finding memory so much as finding out which stimuli is which.
The answer it seems is something called “Declarative Memory” in essence if you “Declare” that you know something, you can find it again. Another way of looking at this, is you don’t know you know something until you tell yourself that you know it. But really what we need is some really enormous index with lots of search terms. Ideally declaring a memory should be as simple as recording an index to it, somehow.
To do this we need two things, a unique address for the data, so it can be found, and a way of tagging it with the search terms we will use to search for it. Now this is where chunking really shines, essentially a chunk is nothing but a list of sub-addresses that make up a unique combination of data.
So all we need to store, is a reference to a particular chunk, and we can retrieve data from the cerebral cortex at will. What we need is a two way map, on one side we need to map the chunk to the index, and on the other side we need to map according to the search terms to find that particular index. Once we have done this, the mere act of searching the map of search terms, can be linked to the cerebral cortex addressed by the chunk.
The main requirement is some connection for the output from the index to the pre-activation stage of attention. It is my contention that the presence of Beta waves at the cerebral cortex, might be an indicator that such a link exists. Especially if the Beta waves are really BSO’s or Beta Synchronous Oscillations.
The basis for this theory is the connections of the Subiculum which might I have thought work for the Declarative memory in much the same way as the thalamus does for the cerebral cortex, allowing separate pre-activation of hippocampal area mapping elements such as the “Place” codes found in CA3.
Once we can reliably access the same memory elements multiple times, the next problem is how do we string the Complicit Commands together into a language for programming? The answer might be skill memory.
In this hypothesis, I stretch Marr’s original concept of the function of the Cerebellum to include not only skill related to action, but skill related to processing as well. One of the problems of concurrency, programming in a parallel environment is the problem of how to simulate a sequence when you are processing in parallel.
Now an actual sequence where one element follows another in lockstep, is pretty difficult to achieve, because of the synchronization that is needed between parallel processes. However a pseudo-sequence can be made from a tree structure where every branch is a decision and the progression is from decision to decision, rather than from step to step.
Given a Tree like structure, all you need for a pseudo-sequence is to have a state machine that moves you from decision to decision. Exactly the function that the attention system was doing in the cerebral cortex. In fact it can even work in much the same manner, pre-activation, and then selection. One might even assume that presence of an ASO or Alpha Synchronous Oscillation.
The idea is that the output which is a sequence of either action triggers in the motor area or selections of data and processing functions in complicit attention is fed back through the thalamus to pre-activate ready for ACC selection.
Why do we need selection you ask? well it works out that this type of parallel pseudo-sequencing is not exclusive, and so instead of getting just one sequence you get the equivalent of a population of sequences, from which you have to choose. The choices in turn have to come from the cerebral cortex. So selection between pseudo-sequences allows us to steer from valid response to valid response without the linear dependency of true sequences.
In order to select between equivalent processes or actions you need some evaluative function which implies a control loop where decisions are made and the results fed back into an evaluation step before more decisions are made. A next step is to link particular pseudo-sequences to particular outcomes, developing a sort of rule-base.
Once we have this rule-base it becomes possible to evaluate sequences of sequences This I believe is where the SMA comes in. In essence it’s role is to evaluate sequences of sequences. An interface beween it and the Cerebellum is likely through the Dorso-Medial PFC This can be seen to be equivalent to a macro-language like Forth, in that we have the pseudo-sequence macros at the cerebellum that can be interpreted as an inner loop and the SMA that can trigger sequences of Macros to form actions and processes, all feeding though the complict attention system to automate our actions.
Now this early level interpreted language, is not directed by will but autonomic, and operating mostly on instinct and emotional drives. But it is directed in the sense that it moves the body and mind closer to it’s goals without even recognizing what those goals are. So we call it intention to separate it from directed programming involving consciousness which we will call Volition.
The primary problem with intention, is that it is relatively un-steered, and un-steerable. It has a tendency to continue doing what it is doing without regard to the comfort of the individual. So what we really need, is a control loop that limits the parameters within which intention can operate.
Now we are getting into the realm of reliability, and the simplest mechanism for reliability is the rewind function. Essentially you stop the program when it gets out of range, and rewind the last function to see if you can get it to stay in range. To do this you can go with either a STACK or a LOGGING approach. In a stack the last processes call code is left on the stack, so rewinding it is as simple as throwing away a stack frame. Logging however is a simple variation on pseudo sequencing. We have not found the equivalent of a stack in the brain but we do know that pseudo sequencing is possible.To implement a rewind log, you report to the log every time you do a command, and you run a parallel mechanism lagging just behind the original command, that evaluates the performance of the process, when it fails, you go back to the log entry just before the process that failed and start again. In the spirit of declarative memory, to know what you have done you need to tell yourself that you have done it, this means that analyzing the log, creates the necessary reflection to run the rewind mechanism but also gives you information about what data is being processed in what manner, and whether it is working or not. This I maintain is the seed of awareness on which our experience is built.
The problem with intention is that it is automatic, and doesn’t have initiative. It can’t choose not to try to rewind so much as choose from among possible alternative scenerios to rewind to. When this automated response fails, a second control loop is needed, that includes the ability to synthesize a new process if necessary to get the program back on track.
Since it has the ability to synthesize new choices, this new control loop which I choose to call volition is the basis of what we call will. The important thing to recognize about this, is that this type of will comes late to the process, and is less the cause of actions and more the reaction to the failure of intention. It is however able to plan, and over-ride intention so as to direct action towards a specific goal. In fact a good percentage of psychological disease comes from this interference with intention and is triggered by cultural assumptions that don’t match the needs of the body. -
Graeme,
This sort of model is v. interesting and useful. Can I suggest you enhance it, by showing how the stages apply to a particular example or two – e.g. a dog passes across vision, (possibly about to come towards you), or a voice suddenly pops into attention from a tv playing in the background – or whatever example you prefer – how do the stages kick in?
-
Ok, I will give it a try, I am not really an example oriented person, but lets start with the dog walking across the street ahead of you.
Before you can recognize that it is a dog, you need to be able to focus on it. Now one of the things that doesn’t come out in my description of the attention phases is the fact that much of this stuff is going on in parallel, with different stimulus periods being more advanced then others. So when the dog first moves into your vision, you do not know it is a dog, nor do you know that it is moving, it is an unexplained stimulus.
Previously started stimulus analysis steps feed back into current step, and tell you that the dog is in front of you, in a particular zone of the environment. This triggers the Orienting instinct which turns your eye towards the dog. The dog is still out of focus so you don’t know it is a dog. However it has briefly caught your attention enough for the orientation instinct to kick in.
Now all that your brain knows is that there is something salient about the area that your eye has turned to. Perhaps it is the motion that attracted your attention, or a difference in color between the dog and the street, or something like that.
At this point the implicit memory kicks in and analyzes the image, gathering information about the nature of the salient zone. You still don’t know it’s a dog, but you are inundated with memories of things that have the same color. You don’t know what the color is, you are just reminded of other things that have that color. Also other areas of the brain might be taking in the shape of the elements in the salient zone, not picking them out, but just detecting the edges. Your brain then converts this image of the salient zone, with its added information into a Functional Cluster, by tagging it with a unique frequency tag in the Gamma Frequency. Meanwhile the emotive centers are evaluating each different salience zone for salience, which will trigger the orientation phase of the next cycle.
Now the signal is translated from a Functional Cluster, to a chunk by passing it through the bottleneck, the result is that now we have a location-centric representation of the functional cluster. This is the end-result of the core-area processing. All this has happened in the first epoch of the MMNn signal.
Now the attention system, edits the chunk, looking for sub-elements that are more salient than others, Here is where the boundary areas are used to define zones of similar color, etc. We have a flurry of activity in the MMNn that results in the peak of the second attention epoch, as each sub-zone is first identified, and then checked for salience. The role of the belt is therefore to find salient sub-zones. For the first time we can pick the dog out from the environment, it has a definite extent, and the color becomes associated with the sub-zone that represents the dog. Now the attention system controls the focus of the eyes and brings the dog into focus.
We still don’t know anything about the dog, just that it is a salient sub-zone of our environment and worth focussing on. Further there are other salient sub-zones that we also focus briefly on, during the eyes natural saccades. Here we get into my random impulse motivation factor, essentially there is no reason other than salience to focus on one element of the environmetn more than another, so essentially we randomly pick high salience zones to focus on, and then move on. As a result, we can see that there is an instinctive tendency to focus more often on eyes, nose and mouths, but otherwise there is little predictability to where the vision alights.
this is important because it keeps the brain from concentrating too early on a particular salient zone, so we can be aware of multiple salient zones. An important part of situational awareness.
Sorry I have to stop there, for now, I am just running out of time for another appointment.
-
To continue with my example, one thing that happens in the belt, is that output from the core, implicit memory, combines with the belt implicit memory, in a connection that might be equivalent to a comparator, promoting only the belt connections that also are linked to the core connections, thus specifying which sub-element is which.
This is where a good understanding of Slip-net comes in handy, because the output from the belt, is fed to the Associative areas, where relation rules are applied to determine what the relationships are between the various different outputs. Douglas Hofsteader developed the “Copycat Program” based on the slip net concept to illustrate this concept where inexact rules are applied, and relationships learned, involving the order of letters in an anagram.
At this point the connection is made between dog and blotch of color, and between strange dogs and biting, so the salience of the dog goes up, as we try to determine if the dog is walking towards us, or not, and whether we are likely to get bit. A connection might also be made between names of dogs we know, and what the dog looks like so we can determine if we know this particular dog. To do that we might want to recruit a specialist module that looks at the dogs face and attempts to connect it to dogs we know.
To do that type of processing we have to stash the dogs facial information into a buffer, and feed it to the specialized function we recruit. This is called Complicit Attention in my model. Which data is triggered and which functions are considered are automatically chosen via instinct/emotional Salience of the now recognized dog concept.
It is my belief that at this point, enough is known for a first rough indexing of the memory. As a result the memory is inserted into our declarative memory index, and the surrounding information is analyzed to determine things like What, When, and Where the memory is. This rough mapping is then applied to the episodic memory.
These index entries can then be used to select memories of that particular dog.
At the same time, an echo of the choice of the data and function, are transmitted to the Cerebellum, where it represents the relationship between existing contexts and the commands, and learns how to make pseudo-sequences from them. Whenever a similar context is triggered, the pseudo-sequence that is related to recognizing the dog is served up, as is similar seuqences such as what to do if the dog is strange, and what to do if the dog is running away from you and is familiar etc.
These pseudo-sequences are used by the thalamus to pre-activate either processing or Action sequences, and instinct/emotion allows us to pick the right one so that we don’t try to kick a freindly dog, or run after a strange dog because we thought it might have broken loose.
The SMA meanwhile is looking at the nature of the sequences we select from the cerebellum, and is trying to string them together into sequences of sequences, much in the same way we would program a computer by stringing together macro programs into a macro language like Forth. When it finds a valid string of sequences, it memorizes it, so that it can act like an outerloop in a forth interpreter, feeding the macros to the cerebellum, to trigger them as if they were a form of location-specified commands.
One of these sequences might be how to meet our own friendly dog, and ruffle its fur so that it knows we care about it. Another might be how to avoid strange dogs that might attack us, by shouting at them and scaring them off.
Now because all this has been autonomous up until this point, we have really had no control over it. That is why we call these automations intention instead of volition.
But consider the case where the automatic responses are wrong. Consider that the dog looks like our friendly dog, but as it gets closer, the face looks subtly wrong, and we no longer recognize it as our friend, or suppose it starts to slaver foam, or look surly and angry. We need an error mechanism that lets us readjust on the fly, to the fact that we were wrong to rush forward reaching for the dog. To do this we need some form of log or stash of information about what sequence we selected, so we can choose not to select that particular sequence the next time we choose, it would be even better if we could rewind to the sequence previous, so we can act as if we didn’t madly rush forward to ruffle the fur of a strange dog. To do this we need a control mechanism that can detect failure of a particular tactic, and change tactics.
This I believe is the reason for Awareness, it gives us a log of what we are doing, that we can analyze in parallel to actually doing it, so that we can stop poorly chosen sequences and restart them with something that is more prudent. But the feedback loop it creates reflects our actions back onto the processing system of the brain, allowing the actions to reflect in our brain as we attempt them. Structurally the log must lag the selection, and the analyses must lag the writing of the log, so there is an appreciable time between selection and awareness that a selection has been made.
Now we have an opportunity to override the intention, with a better strategy.
To do that we need to have developed a strategy, and used the strategy to link macros together into a sequence according to the strategy. Along with this we can select our strategy to fulfill a particular goal. But the programming generator is not error free, what if our strategy picks the wrong tactics or parameters, and falls outside our comfort zone?It would be nice, to compare a new program against a model of our comfort zones, to predict if it would fall outside the comfort zone. In fact we might want to compare the tactics in the intention sequences as they are expressed to determine if they are crowding the comfort zone, and adjust them if they are.
this is where the self-model comes into play, used as a model of our comfort zones it can be used to link tactics with our models approval, or a model of the program can be compared against a model of the self, and we can see that it will possibly push our comfort zone at some point, and thus its strategy can be adjusted to pick less radical tactics. A “Self” signal indicating that the tactical choice has passed the model would help us choose between possible tactical sequences.
To determine whether the strategy is correct however we should use the same sort of logging construct to monitor strategic choices, and this forms the reflective effect I call Consciousness. Because we choose our strategies, the automations we form at this level are called volition. Perhaps we recognize the dog, but it is not an especially freindly dog, and we know that it has slipped it’s leash. We might want to carefully approach it, with the idea of capturing it and taking it back to its owner. However on approaching the dog we might find it surly, or suspicious and have to give up capturing it. Changing our strategy to telling the owner where we saw it last and in which direction it was moving.
Does that example help?
-
Graeme,
Many thanks indeed – extremely helpful. Work on it some more – with more examples – and you may have enough for a book. Such a model is v valuable in getting us to think about the stages of attention & awareness/decision.
Odd comments. I would think at an early stage, the processing must consider where the dog is moving/going – and therefore try to predict what may be about to happen. [I recommend Jeff Hawkins On Intelligence – who is v. strong on the predictive function – just in case you haven’t read]
It is fascinating to speculate as you then begin to do, about how memory will be used not just as you well analyse to identify the shape and form of the dog, but to compare movie sequences of actual and possible movement of the dog.
How do you feel about one of the big controversies here – cog sci & esp. AI are wedded to the notion of the remembering mind, or any mind, searching systematically through databases of options – I, and AFAICT Hawkin, think that is way too elaborate – the brain must be able to compare shapes by a process of mapping – & by somehow picking out similar shapes from memory with extreme rapidity – in perhaps just a few moves, rather than the thousands-to-millions of modern computers.
Re awareness, clearly the main function, I suggest, is executive decision – for deciding what to do about the dog here, because there will be more or less equally attractive options. “Consciousness is there for when you don’t know what to do..” as s.o. put it. Indeed I don’t think your model makes full sense if it doesn’t end in exec. decision. [Error correction is peripheral or occasional – decision is continuous].
So I found the first stages of your model esp. interesting – the vague shape slowly being identified – but I think the later stages – the inference about what is happening, [what the dog is likely to do],and what can be done about it – could do with further development.
And it was interesting to see you approving of Hofstadter. I think COPYCAT is purely logical analogy and useless for creative analogy, which is the real challenge for AI. But I may have missed some other aspect of that prog.
If you should work on the model more, I would be delighted to see it – and I think it would be even more valuable – but it must use examples. They make the whole thing come alive.
-
M. Tintner said “I would think that early in the processing, the brain would deal with whether the dog is moving.”
I think that that is correct. In my third epoch of attention the brain detects associated information about the location of the dog. Probably whether or not it is moving, and in which direction. Up until that point, while the detectors could detect location, and movement, and direction of movement, they could not be associated with the dog. They could however be linked with the salience zone, and so, emotionally we would be set up to expect something from that location to be moving towards or away from us.
It is an important distinction I think that first you have to detect something, and then you can integrate information about the relationships that something has with other things in the environment.
On the subject of search vs. rapid detection, I have to bring in my Implicit Memory Model. Essentially what I am talking about is a massive rapid detect and react network that outputs parallel and redundant data, about any specific salience zone. Now as I have tried to state again and again, this produces a data-rich, but organizationally poor cloud of data. The output of the Core areas of the brain is sensory Modular clouds of data. In the second epoch, this data is mapped to modular sub-zones, and by analyzing the salience of the sub-zones, patterns of greater or lesser salience are detected, requiring almost constant attention management to address, and compare the different sub-zones with existing known objects. In other words over about 160 milliseconds any functional cluster can be reduced to objects.
Now the main difference between this and the A.I. search, is one of architecture, essentially in a sequential search you put the search term into the search engine and it tests each item to see if it is the right one.
In this search you put the search term into the search engine, and all the possible objects, test the search term in parallel to see if they are similar. The result is a data cloud that consists of all the similar objects in the database. There is only a very subtle difference in the architecture of the tissues between the core and the belt, but that architecture is significant in that it allows the belt to compare the signal from the core with the signal from specific objects. In both cases we are talking about a form of implicit memory, however in the belts case it is the interface between the implicit memory from the core, the implicit memory from the belt itself, and the mechanism that creates Neural Groups, that triggers the choice of objects.On executive function. The role of the ACC is to select by suppression, at almost every step in the process, the advancement of data from one stage of processing to another depends on the ACC selecting which data will move to the next step. In this system executive function is virtually ubiquitous to the operation of the system. Executive function is not the last thing, but nearly the first thing that is needed to explain attention. However you might have a point, my model is not meant to capture human consciousness, I am merely trying to capture animal consciousness, and so I am distinctly light on the explanations for deliberation, thought, and language which I think come after my model, although deliberation to some lesser extent might be found in simian brains.
I think that you are right, I need to do more work on the last few stages, as I am not sure what the ACC has to do with them, or whether they are a role of a different area of the PFC.
On building more of a background in examples, well like I said I am not example oriented so my first attempt was to model the mechanism and see how it might work. I know I am good at that, because it involves abstraction and mechanical aptitudes where I shine. Building examples to explain I think I will find more difficult, but it would be necessary I suppose if I was to want to publish a book on the subject.
-
Graeme,
Thanks. I would strongly urge you to incorporate examples routinely. They massively improve the theoretical thinking too.
You might like to check out Hawkin – he believes, I think, that the object has to be moving in order to recognize it, (wh. I guess might include that the animal viewer is moving around/towards the object). [I personally am strongly incllned to think that the brain is first and foremost a movie / moving image processor – & that we have to overcome our literary/bookish disposition to think of it in terms primarily of processing still images].
Re your cloud – or could it be “web”? – of memory associations, used to recognize a form, roughly how many would you think the brain can activate in a split second, to recognize that dog form? (Wild guesses allowed). Esp. allowing for the speed of neuronal transmission.
-
Lets not talk with imprecise terms about time. The “Split second” is too imprecise, do we want to discuss epochs of about 70-80 milliseconds, or do we want to discuss memory cycles of considerably less.
Further while Web would work, Webs tend to be two dimensional in my experience, and so they limit the complexity of the system. One reason I use cloud, is because it is amorphous and unformed and therefore does not impose an implied dimensionality on the concept.
Think of my work as being not only accepting of imprecision, but demanding of it. I was recently watching a video about a Blue-Gene experiment that had about 10,000 neurons. They went to great lengths to be statistically precise about the structure of the column. In essence they were using a massive super computer, and asking for more power just to implement a single column.
I can’t afford that approach, so instead I have to base my work on imprecise methods. The researchers in the video were talking years before they could even begin to approximate an idea of how their column actually worked, Years of research by top scientists. I already have a model that might work, but it is only once they do the research that we will be able to improve it, to make it more accurate.
Instead of runaway precision, I have chosen to study tissue level detail, and abstract from it, a plausible mechanism. They think that they are the first people to think of mapping the brain at the dendrite level, instead of at the soma/spike level. Perhaps at their level of science, they are, but at my admittedly smaller accuracy level, Dendritic complexity was already recognized as an indication of processing complexity.
They are dealing with hundreds of different types of cells with different electrical characteristics, based on over 200 different ion channel possibilities, I have abstracted them into three main types of cells by role, mostly characterized by their dendrite and axon formation.
Perhaps my model will be found to be oversimplified, I expect that such is the case, but then, I am not trying to design a micro-chip, and they are.
On the other hand, their standardized micro-chip, will not be well suited to a role in the core-belt transfer architecture. They will either have to have two microchips to achieve that action, or they will fail to capture the function of even the first three epochs of attention.
On the question of how many elements can be recognized at the same time, the problem is complex, or simple depending on how you look at it. I could over-specify and try to calculate how many cells there are in each column like the guys working on Blue Gene, and then specify how many cells there are in the cerebral cortex, and then try and take a statistical sampling of how many fire at the same time, or I could just point out that the question is mute, since with massive parallelism potentially every cell in the cerebral cortex could fire in parallel. On average they wont, and for good reason the blood demand would outstrip the supply, but how that is regulated at the neural level is beyond me. Even small areas of the brain that are overactive, in diseases such as epilepsy, have the effect of causing convulsions, so I would suspect that there is a distinct limit to how many cells can fire in a short period of time, without pathological symptoms, but the fact that people get epilepsy suggest that it is a soft limit, and that therefore we can only approximate it.
So in my imprecise way, I will limit myself to saying “A LOT!” even though the extent of the Belt area is limited we have to take into account that there are multiple belt areas that may be linked via the original functional cluster in a cross-modal pattern. The association areas are also limited although not as limited as even the belt, and so the associations that allow the recognition to link to a specific conceptual classification are also relatively large, although they tend to cluster in smaller areas within the larger association areas, that specialize in certain types of associations.
For instance some association areas tend to associate the “What” of a storage element, while others the “When” and “Where”. I am not sure of where the “When” aspect is linked but suspect that in slip-net fashion, it is merely a measure of “Where” the entry is with respect to time. It really doesn’t seem all that expensive to determine precedence in a sequence, such as we know is created in the Parietal Lobes, and fed to the hippocampus in order to determine place, the question is where is the processing done, and I can’t answer that question with my current level of knowledge.
-
