Artificial Consciousness/Neural Correlates/Functional Models/Attention

The Attention System

edit

I Hypothesize the existence of at least 6 separate different types of attention. There may be more, and scientists could classify them and lump a number of different types together into a single more integrated model In fact, I would do that myself, if the system were not more complex than a smaller number would properly deal with. Simply put, there are the following distinct Attention types:

  1. Instinctive Attention With Dorso-Lateral Suppression
  2. Implicit Attention with Latero-Ventral PFC Involvement and DorsoLateral Suppression
  3. Explicit Attention with Latero-Ventral PFC Involvement and DorsoLateral Suppression
  4. Intentive Attention with Latero-Ventral, Orbitofrontal and DorsoLateral Involvement
  5. Volitional Attention with Latero-Ventral, Orbitofrontal and Dorso-Lateral Involvement
  6. Subicular Attention at the Entorhinal Cortex (Probably for Episodal Memory)

Crick has suggested that Attention is like a Spotlight in a dark room hi-lighting only the memory elements it is pointed at. However for some of these attention systems this characterization is not germane. However it is a powerful metaphor and should not be put aside lightly especially since it fits so closely with the folk-psychology concept of attention as being directed at a specific memory.

The problem with the Spotlight analogy, is simply it implies direction before attention, which brings back with it the problem of a homunculus. If we want to get around the homunculus problem we have to move away from the spot-light metaphor. I believe that properly formulated this hypothesis does just that.

Instinctive Attention with Dorso-Lateral Suppression

edit

Part of the problem with Folk Psychology, is that it wants to impose direction too soon. It can be shown that intentive actions take a certain amount of time, and that Volitional Actions take longer. In fact this is why Libett's work is so controversial, he has suggested that a Volitional act averages 500 milliseconds. The body does a lot in 500 milliseconds, which supports the contention that much of what it does, does not require volition. However, Folk Psychology demands direction which can only be volitional, in order to work the attention system.

Whether you agree with Libett's timing constraints on volition or not, is really immaterial, some Attentional states, pre-exist volition, and some even pre-exist intention. In some cases we can say that the body is reacting before the signal has fully been analyzed by the brain. Now if you think about it, such a reaction is, in fact evolutionarily sound, in that Quick Responses to danger allow more time for processing once the body is already moving.

Specifically what we are talking about are instinctive reactions like jumping or turning in the direction of a critical stimulus such as glass breaking. These reactions are tribbered instinctively before the brain is capable of picking out where specifically the stimulus is coming from, as a result they are dull responses, no matter how sharply we move to meet them. There is no intelligence in always turning towards the sound of glass breaking, so the mechanism of the brain is set to suppress these reactions where they are clearly un-called for. Suppressing them requires a recognition that they do not offer as much danger as something else, or for that matter, that they should be suppressed which comes later in the Attention System after some processing has occurred.

As such, the actual instinctive reactions are mostly unconditioned reflexes, and learned reflexes, not anything that would require thought.

It is my theory that these reactions are triggered in the Basal Ganglia, and suppressed by the Dorso-Lateral Prefrontal Cortex, as part of either intention or Volition.

Implicit Attention with Latero-Ventral Involvement and Dorso-Lateral suppression

edit

In his book The Mind doesn't Work that Way! Jerry Fodor makes the point that phenomenal systems have phenomenal outputs, and that it is impossible for a phenomenal (Neural Network) to be addressed by individual memory elements, if only because the locations of the elements are a function of the network as a whole, although they may be localized within the network depending on its design. There is nothing in our understanding of Dr. Edelman's work that suggests that we should be able to just reach out and address a specific memory even at the Neural Group Level. In fact he makes the point, in The Remembered Present that neural groups are interchangeable. That any coding system to address them must be in fact Degenerate and unique to the individual. But even if we had such a coding system, the actual locations that are related to a specific object are spread across the sensory modalities by which we know that object. And because there are few interconnections between lobes, those sensory modalities are several from each other.

The spotlight metaphore needs a lot of help in this environment, since there is no target for it to point to. If you can't use a place-code address to find something in a phenomenal memory system, then, Theoretically the only way you can address specific memories is to address them by content. This is what Marr's model of the Cerebral Cortex claims the cortex is doing. The problem with that, is that you get volunteer memory instead of directed or demand memory. What happens is you put a stimulus onto the input bus, and any content sensitive elements in the network respond by dumping their contents onto the network, if they recognize it as similar to what they contain. this is a distributed memory application, and doesn't look at all like a spotlight. Redundancy is required.

This is why implicit memory is so rich, in associations. Essentially it defines a field of self-reinforcing data. We call this field a Quale. The main problem with Qualia, is simply that because of their phenomenal nature they can't be subdivided. This puts a distinct limitation on the nature of memory in the implicit memory system, We can filter the data, but we can't break it down into distinct elements. Now don't get me wrong, it is not that we are not analyzing it down into distinct elements, but that the analysis results can't be separated from the recognition results, because they are all balled together in this Quale.

We can, however use the inputs from the quale, to define linked data fields, that are subfields of the original datafields. But this must be done with filter techniques not with selection at the memory element level. We have already suggested that the early results of the sensory system are somehow getting analyzed by the Basal Ganglia, so it is my hypothesis, that they are also being roughly grouped by the basal ganglia into clusters that related to specific fields of view, to assist in the selection of which direction to turn, to orient on a sound, or a flash of light.

Dr. Edelman has called these Functional Clusters, and suggested that they solve the Binding problem which is how does an essentially stove-pipe sensory system bind different sensory modalities into a single entity? They are bound because the Basal Ganglia Binds them. But the way we can detect they are bound is by the GSO signals that make them synchronize at about 40 hertz. Unfortunately for this hypothesis there is no reciprical GSO signal that has been detected comming from the Basal Ganglia, which doesn't have the required mechanisms to impart such a frequency as far as I know.

This means that some other mechanism is probably involved, and the likely culprit is the Thalamus which is part of the Reticular Activation System, which is thought to induce brain waves, and is also involved in the Bottom-up Attention system which is implicated in pre-activation of neurons as part of the attention system. What I think is happening is that the Thalamus which has been linked to the GSO signals because it resonates at their frequencies, is involved in distribution of the signals. The main problem I haven't yet solved is how the Basal Ganglia tells the thalamus which cerebral cortex areas to pre-activate with each GSO frequency.

Once the cortex is pre-activated and the frequency is imposed on the otherwise unchanged signals coming from the 2nd. and 3rd. Layers of the cerebral cortex, We can filter for specific Qualia, simply by only allowing one frequency to pass a filter, and basing our selection on the GSO frequencies generated by the thalamus. What we are talking about is not Cricks spotlight but more of a flood-light that illuminates broad areas of memory. The Latero-Ventral Prefrontal Cortex is probably involved in this filtration. Thus setting the prefrontal cortex into a role as Selector, of Activated signals pre-activated by the bottom-up attention system. This is much less than the role suggested for the top-down attention system, but we are only working with a portion of it, at this stage.

Explicit Attention with Latero-Ventral Involvement and Dorso-Lateral suppression

edit

Now the problem with this filtered attention, is that we quite literally can hardly do anything with it. We need that spotlight, but we can't get it, because the output is a Quale, which is still an indigestible field of data. All we can do with it, is project it onto a processing center like the Basal Ganglia, and instinctively react to it. Evolution wasn't going to allow that pass to happen for long, although it might have been enough for whole species at some point in evolution, so the problem was how to convert from this undigestable content addressable form, into a more useful and digestable form, probably with some sort of demand memory so we could pick out the actual memory elements from the Quale. Jerry Fodor, pointed towards the Column interface. If there was some locality to information within the network, something he wasn't convinced was possible, then there would also be an architecture that allowed memory to be addressed by that locality.

David LaBerge, however has linked Bottom-up Attention to the 5th layer Pyramidal Neurons in the Cerebral Cortex, and has told us they are organized into Mini-Columns within the Column architecture of Jerry Fodor. If you have been following all my arguments up until this point you will know that this means that the Bottom-up Attention system might be used to address the Cerebral Cortex at somewhere around the Neural Group resolution. Now Dr. Edelman has assured us that neural groups are interchangeable, so we can't select a specific memory using this mechanism until we develop some kind of index, and we know that phenomenal systems are self-adjusting, and thus may sometimes change the Neural Group a specific memory belongs to, but we would have the ability to address those memories if we could find them. Our memory system would have to respond to loss of an index to a specific memory and that might be the role that the Tip of the Tongue Feeling is about.

So the main problem is to convert the memory from an implicit to an explicit form and to index it somehow. Once we have done that, we have a demand memory. We can sample the memory by individual memories, and Process that, and base our actions on the results. That sounds like the spotlight Crick was talking about!. Of course it isn't quite that simple, We have an addressing technique but no way to tell us what memory is in what location within the now demand addressable memory system, and worse than that, philsophers tell us that our output is still a Quale! That means we have to designate the memory we want, blindly until we get our index up and running.

Well lets look at what Annette Karmiloff-Smith has to tell us about memory in her book Beyond Modularity She describes a process called representational redescription that she claims involves the conversion of implicit memory to explicit memory, and is necessary for learning. Essentially what she tells us is that memories have to go through a multi-phase process before they are fully learned. Worse this process can take Years, which is why it takes 6 years for most kids to get fluent enough in the spoken language that they can go to school. Apparently you can't rush it, because memory has to go through each phase to be ready for the next.

Well with techniques that I have started to develop and describe in my book Dataminnig Intuition you may be able to take a few shortcuts, but the process is more or less set in stone. The memory moves on in it's inimicable fashion no matter what we do.

While it might be nice to do it a more sophisticated way, one way we can build our index is using brute force methods, of sampling memories and trying to figure out which are more important to index. A slightly more sophisticated game plan might be to work back from existing functional clusters, and use a heuristic to reduce the cost of finding each memory. Whatever the technique it is obvious that quite a bit of processing is going to be involved in discovery of the locations of specific memory elements. But the heuristic search from existing Functional clusters at least has the added grace of being related to our survival, since it deals with the nature of memory as it becomes important. To do it however we need a data element that has in the past been called a Clump, to hold list of the individual addresss of each mini-column in the functional cluster. This means our Representational Redescription between Implicit and Explicit forms of memory is the finding of the (in Mini-column addresses) contents of the clump somehow.

I will leave out the theoretical method of doing that gathering of mini-column addresses that I describe in How the Memory Might Work because it is not really germane to the problem of attention, and might be too speculative for this venue.

Intentive Attention with Latero-Ventral, OrbitoFrontal, and DorsoLateral Involvement

edit

Following the strategy I laid out where the brain uses an existing Functional Cluster converted to a clump, as it's base, and then uses a heuristic to reduce the search time for discovery of memory elements within the clump, we need the ability to project the clump back onto the cerebral cortex, and recover a new Quale. This approach called Rehearsal, has been proven to be important to the size of the short-term memory. Supporting the contention that the body actually does rehearse clumps it has already converted into explicit memory.

To understand how this works we need to understand that the clump while it selects all the mini-columns involved in the original memory is really only an approximation similar to the approximations made to digitize music. To recover the original quale, we need to present it to the cerebral cortex as a stimulus pattern and collect the content addressable memories interaction with the stimulus. So our output is still a Quale. Ok, no problem. Or is it? How do we select from among the mini-column addresses a subset, that we can evaluate for saliency?

The answer as near as we can tell, is that we need to have intention, the ability to decide which mini-column addresses to try for which functional cluster. The problem is, at least at first, there is no mechanism by which we can know which mini-columns might be best, at least until we have some sort of cursory index up and running. Now we get into a bit of speculation but I think baby activities near birth support this contention, the answer is to use a random impulse as the selection mechanism, and then as we get more information and can make better choices, replace the random impulse with guided selection.

The trick is to project the Quale onto a processing field similar to the basal ganglia that triggered our early instinctual reactions, and clustered our original quale so it could be filtered, and determine which combinations of unit fields have saliency. This feeds back into the prefrontal cortex an emotional signal called the feeling of correctness. The better the saliency of a particular combination of mini-column addresses the stronger the feeling of correctness. So at first the baby randomly samples their functional clusters, and builds up experience finding the most salient combinations of mini-column addresses, and then it uses that to overtake the random impulse and direct the search to use the best strategies it has found. The result should be a rapid increase in the quality of the search for memory elements. This rapid increase is created by selection from among strategies including random impulse.

As it makes decisions as to which strategy to use, the brain also builds a model of the strategies it likes to use. Called a Self-Image, this model is monitored by the orbitofrontal PFC and the decisions become linked to a feeling of self, associated with an agreement between the model and the choices made. This is the first hint of Agency. The process by which it selects a specific mini-column address to include or exclude out of the clump, for a specific memory is problematic, if only because all we have is a random signal for the basis of our selection at least to start, and the full clump of mini-column addresses. For this part of the hypothesis I have to go to David LaBerges Triangular Theory of Attention. What he described was a linkage between the PFC and the Thalamus that since his work has been traced to a connection to the Nucleus Accumbens and from there to the Nucleus Reticularis Thalami, a sort of Layer wrapping around the thalamus, made up of inhibiting neurons. What this seems to suggest is that at a certain point in the rehearsal of a specific clump, the prefrontal cortex can select which mini-column addresses to inhibit, thus reducing the size of the clump to some subset of the original clump. It may be possible to do this type of thing multiple times in the rehearsal process, thus searching multiple random selections for the most salient one in the same rehearsal step. The Nucleus Accumbens might act as a sort of Latching buffer stabilizing the address before it latches it through to the output that influences the nucleus reticularis thalami. The area of the Prefrontal Cortex that makes the decisions between the strategies is probably the Dorso-Lateral PFC, using the same mechanism that it uses to suppress earlier forms of attention.

Volitional Attention with Latero-Venteral, Orbitofrontal, and Dorsolateral involvement

edit

Although we are not sure, at this point what the main difference between volitional and intention are at the attention system, we are fairly sure that the main difference is the requirement for greater processing during the decision making. What we think is happening, is that a sort of model of the environment, is being compared with our proposed actions to predict what their outcome might be, and then this model of the outcome is used by the dorsolateral PFC to select from among different strategies, the ones most likely to result in favorable outcomes. In order to make the success or failure of a particular strategy feed back into the model, there needs to be feedback from the activity, to be evaluated in near real-time for success or failure of the strategy.

One interpretation of the feedback is that it is the experiencing of this reflexive feedback that we call awareness. Because we are aware of our actions, although not of the processing going on below them, we call decisions made during this type of processing Volitional. An interesting thing to note is that the signal that is measured by Libett and his contemporaries, is a sort of suppression of output signal that is negated when we decide to do a volitional act. It has been jokingly said that obviously at this stage Won't Power is more important than Willpower.

Subicular Attention at the Entorhinal Cortex

edit

While this interpretation might be less than valid, documentation that the subiculum feeds into the 5th layer of the Entorhinal Cortex, reminded me of the bottom-up attention system that influences the 5th layer of the Cerebral Cortex. If there is any functional overlap, the subiculum might be responsible for pre-activation of the Entorhinal Cortex Mini-column Addressing in much the same way that the Thalamus is responsible for similar pre-activation in the cerebral cortex. If so, the link between the subiculum and the DorsoLateral PFC, and the Entorhinal Cortex and the DorsoLateral PFC, might create an opportunity for data from the hippocampal area to feed back into the working memory.

We expect some such mechanism, if only because we can see our minds using episodal data. Having control over both the subiculum and the entorhinal cortex at the DorsoLateral PFC would allow the DorsoLateral PFC to select from among memories in the Hippocampal area in much the same way that the Thalamus allows us to select from among memories in the Cerebral Cortex.