Monday, 28 May 2012

An Evolutionary Model of the Brain's Internal Language

The purpose of this post is to suggest that the brain has a simple internal language for storing and using memories of objects and events, that this language is common to animals and humans, and as the capacity of the brain expands the language has sufficient power to support human intelligence. In particular it identifies two areas where changes in the learning/teaching mechanisms could have led to a tripping point where it became economical (in evolutionary terms) to develop a larger brain. It suggests that much, if not all, of what we consider intelligence is the result of cultural factors.

The following topics are addressed
  1. The Black Hole in Brain Research – bridging the gap between the neuron level and intelligence.
  2. Evolutionary Pressures on Brain Size – bigger is not necessarily better
  3. The Simple Brain Model – the minimum requirements of a learning animal – introducing “Memodes”
  4. A More formal Definition of a Memode
  5. Introducing CODIL - the history and references to this blue sky research project
  6. Why CODIL is relevant - and some factors to be considered
  7. Evolving More Brain Power - instructing and programming cultural information
  8. What Next? - Ensuring the ideas get followed up

1. The Black Hole in Brain Research

An unbelievable amount of very detailed research has been carried out world wide on the inner workings of the brain, into the nature of natural language, how children learn, intelligence in animals and humans, psychology and brain disorders. However if one stands back and takes an overview there is no unified model which allows one to follow a step by step route between the electrical impulses at the neural level and, for example, how a child learns to talk or why we enjoy Shakespeare. The purpose of this discussion document is to go back to first principles and suggest a framework which, with further research could help to bridge this gap.

The approach taken is to start by assuming (1) that there are a few basic information processing facilities which any animal capable of modifying their behaviour as a result of learning must have, and (2) that the human brain works in basically the same way – with no major changes – the differences being mainly related to increases in size and the number of links. In effect the approach suggests that a human is just a normal mammal except that, like the giraffe and its neck, or the elephant and its trunk, it has significantly expanded the way one particular organ, the brain, is used.

Some of these ideas have been discussed earlier as The Black Hole in the Brain.

2. Evolutionary Pressure on Brain Size

It is very easy to assume that a bigger brain is always better but in evolutionary terms this is not the case. Running a brain is an additional (and non-trivial) expense and the more resources expended on the brain the less resources are left for other needs and I have discussed some of the factors in Evolutionary Factors starting on the African Plains.

The problem is that running and “loading” the brain with information has a cost, with everything learnt being lost at death. An animal with a bigger brain will need to learn more to “fill it” with knowledge and this will take time, possibly extending the childhood period. If there is any form of “instruction by example” by parents this also has resource implications for the parents. For herd animals it is advantageous for the young animals to play “follow my leader” with the older animals. A young animal does not need to know about predators if there are always older, more experienced animals there (not necessarily of the same species) as long as they run when the herd runs. This reduces the need for larger brains because vital information is automatically shared with minimal “teaching effort” by the older animals.

This is developed further in social groups of primates and elephants where there is a positive grandmother effect where the older, more experienced animals can support the younger adults, such as a mother in labour for the first time. Because information is passed positively between generations, knowledge is not completely lost, so there could be advantages of a more “intelligent” brain.
A critical factor is how much information can be economically passed between generations – and with the coming of human language (and later writing) there will be a need to evolve a means of faster learning to reduce the learning and teaching times, and possibly a need for passing on learning techniques. This is a factor which must be addressed in the current model.

3. The Simple Brain Model

An animal which is capable of learning must have a facility which will remember “objects” and recognise them when they recur. For instance a chimpanzee must be able to recognise a banana, and it would be useful if a mouse could recognise a cat. At this stage we will not consider how this information is encoded in the brain but will simply refer to the unit which stores and subsequently recognises real world items a memode.

The job of the sense organs, and the associated parts of the brain is to generate memodes. In a new born – with an “empty” brain these memodes will be stored – and if the associated pattern repeats the memode will be reinforced while the ones which do not repeat will be lost. This will build up a language of “objects” recognised - and this reflects what happens when human babies learn the phonemes that make up the language of their parents.

In the real world the animal will be aware of memodes relating to more than one object – and any effective learning will have to relate how objects interact. In the current simple model the current context consists of the most recently activated memodes, which will record what the animal is currently seeing, smelling, etc. At this stage the brain will need to link together memodes which are active at the same time, using a similar learning pattern for the links – with common links being reinforced and uncommon links being lost. Once this has happened the network of linked memodes can start making predictions.

To take a simple example where the linked memodes in a chimpanzee represents HUNGER – BANANA – EAT. When the body activates the HUNGER memode the link, in effect, interrogates the BANANA memode to see if it is true – asking “are you currently seeing a banana?” If the answer is yes the link to the HUNGER memode is triggered – and this in turn triggers the action of eating. If no banana is currently being seen the BANANA memode will have been generated by the vision section of the brain so could start a visual search for bananas. Such referrals back to the originating sense areas of the brain could explain why, when thinking about something, some people have a strong sense of seeing, hearing or smelling what they are thinking about – and also relate to the imagery of dreams.

How the memodes are linked

Let us consider that a mouse might have memodes representing the SIGHT-OF-A-CAT, the SMELL-OF-A-CAT and the SOUND-OF-A-CAT and while often these would occur together they could also be actively singly. If they link together to form a new memode CAT this could be “true” if any one of them was “true”, or is a combination of them was “true”. When we get to thinking about humans and language one can easily add in WORD-FOR-A-CAT as another way of identifying a cat.

But the SIGHT-OF-A-CAT memode might similarly link to memodes representing the EYE-OF-A-CAT, the WHISKERS-OF-A-CAT, the SLINKY-MOVEMENT-OF-A-CAT, etc. and these in turn could have links to memodes concerned with even smaller details. This demonstrates an important feature of the model. The memode network is recursive, and, assuming that only some of the links at any stage have to be true, provides a form of inexact (or fuzzy) matching. In humans, and I would assume also animals, one is only aware to the highest level memodes. For instance as a bird watcher I may instantly recognise one bird while being unconscious of the individual features – but if there is a problem I will look closely to see if I can identify an detail which will distinguish between several possibilities.

A similar approach can be used to classify hierarchies of different but related objects. For instance the ANIMAL memode would link to CAT, DOG, ELEPHANT, etc, and the CAT memode would linked to individual cats such as TIDDLES, GINGER, THOMAS, etc. Memodes would be in several hierarchies and CAT might be in PET, BIRD_KILLER, etc. as well as ANIMAL. The extent to how far animals can take this degree of generalization is unclear to me, but it is only a variant of the previous example involving the SIGHT, SMALL, and SOUND-OF-A-CAT.

There is one other way in which memodes can be linked, and this is when very different things are linked in a single event. A pet dog will link MASTER-STANDING, LEAD-IN_HAND and WALK, and leap up, tail wagging in appreciation, when its master stands up and picks up the lead. Some more intelligent dogs might even attempt to make LEAD-IN-HAND true by picking up the lead and taking it to the master!

In many cases a linked series of memodes such as these can be likened to a program statement which interrogates the currently active memodes.


4. A more formal definition of a Memode

A memode is the active unit of information in the brain. It represents an object, abstract concept or event which needs to be remembered. While a memode might directly hold a low level piece of information, it will often consist of links to other memodes. Any memode can be highly active in the current context – which means that it is at the focus of the brains attention. Otherwise, If triggered by another memode it can compare itself with the memodes in the current context and recognise if it matches. It can then either signal “true” back to the triggering memode, join the current context, or trigger other nemodes.

Some memodes will be linked with the senses, while some will be linked to the brain “outputs” - which will be triggered if the memode enters the current context. In addition each memode will have some kind of learning function associated with it, so that frequently used memodes have a higher priority, while unused memodes eventually get forgotten.

In the brain a memode could be a single neuron or a linked group of neurons. 

This simple model represents what can be considered a minimum for any animal which can learn to recognise a reasonable number of real world entities and make simple decisions. The key question to be asked is how far the human brain works in this way, taking for granted that it has greater capacity in terms of the number of neurons and the number of interconnections. If we start from the bottom up and assume that there are no fundamental differences (apart from size factors) between the animal model, described above, we will either end up showing that there are no fundamental differences or we will be in a better position to understand what they are. Studying the model may also suggest where the main evolutionary pressures were.

5. Introducing CODIL

Fortunately there was some research some years ago which used a similar model and while there were differences, which will be discussed later, the experimental research carried out then showed that such a simple approach could pack a lot of processing punch.

CODIL started in the 1960s as a piece of blue sky research into alternative information processing architectures which aimed to avoid some of the problems inherent in the conventional stored program computer approach (see The Limitations of the Stored Program Computer). The underlying idea arose from an unconventional approach to handling a very large commercial task (Working with LEO III Computers at SMBP) and further research was actively supported by two of the pioneers of commercial computing in the UK, David Caminer and John Pinkerton, The hope was to provided a “white box” processor, which would allow humans work with it to handle open ended tasks which could not easily be predefined in advance. CODIL was, in effect, the human friendly symbolic assembly language for the proposed new architecture. The basic principles were clearly demonstrated in a series of publications ending in a schools package which demonstrated its ability to handle a wide range of artificial intelligence and other tasks which received excellent reviews, followed by a detailed paper in the top ranking Computer Journal . Unfortunately the project was aborted before the final paper appeared in print for health related reasons.

While CODIL was conceived in terms a modified microcode “instruction set” within the central computer processor, the research was carried out with four different software simulations, and each was tested in different ways. The first was simply a test bed to demonstrate that the idea would work. The second was designed to test out a number of commercial data processing tasks on a small scale, and was used to set up a research medical data base and a historical data base. There was also a major excursion into artificial intelligence with TANTALIZE. In 1980 online access became easier and the emphasis switched to using CODIL to support teaching packages and to demonstrate how it might support an interactive publishing network (conceptually a text ancestor of the world wide web). For various reasons the early software was not very portable and with the coming of the the first small personal computers it was decided to demonstrate that the idea was simple enough to fit into a BBC Micro (effective memory 25Kbytes!!!) and the result was MicroCODIL which prove logically far more powerful than its predecessors and which attracted rave reviews from a wide variety of magazines. Unfortunately at the time the MicroCODIL software was being produced there was a suicide that was a major factor leading to the closure of the project in 1988. The final publication appeared in the Computer Journal in 1990.

6. Why CODIL is relevant

While CODIL was implemented on a stored program computer conceptually it was very different as is explained in the post How Many Trucks will your Helicopter Pull:

Computers were originally designed for highly repetitive numerical tasks involving well designed pre-defined algorithms. The underlying mechanisms were never designed to be human-friendly - and to most people the computer in effectively a "magic black box." The more complex the task the more effort has to be put into pre-defining the algorithms and it would be interesting to know how many billions of man years (including time spent up blind alleys and on training) have gone into creating the computer-dependent world we now live it. While the benefits may be enormous, getting there has not been cheap.

CODIL starts from the opposite end of the spectrum. It effectively starts on any task from a position of ignorance - with nothing pre-defined. It does not formally break information down in to program and data, and there are no global rules. In effect it works by comparing information which it has just received with copies of information it has received in the past and uses this make simple deductions, nothing more. The fact this works may surprise many familiar with the limitations of the stored program computer, but many tests have shown a wide variety of situations where it demonstrably works. The whole is set up so that the human who is inputting the information can see into the "white box" so that he can understand how the information is being handled.

The important thing is that CODIL was an attempt to provide a flexible information tool which will help humans with complex open-ended logical tasks which cannot be fully pre-defined in advance, and to do this it had to model some aspects of how the human brain handles such problems. The detailed working are described in the papers, cited earlier, but the key features are a central “decison making unit” which scans linked lists of “items”, comparing them with a list of items called the “Facts” which provides a working short term memory. The principal difference in the underlying approach is that in CODIL there is a single central processing unit while in the memode model the decision making is distributed over the network.

There are many aspects of the CODIL research which could be discussed in considerable detail at this stage to assess their relevance to the memode model but to keep this post within a reasonable length I have selected three where I feel that further work is required - and am very happy to provided information on others on request.

CODIL was developed as a tool

The first thing to consider when assessing the relevance of the result of the CODIL experiments is that CODIL was designed to be a practical tool which could be used alongside conventional computer systems. It was therefore taken for granted that, for example, any usable package should be able to carry out a comprehensive range of mathematical calculations. It should also be able to handle keyboard input and display what it is doing on a screen, or route it to a printer. For this reason the experimental software packages contained libraries of support routines which are not relevant to the brain model. In the case of the package MicroCODIL about 95% of the software was there to provide a dedicated text editing facility and about a dozen different windows to show the user what the system was doing. (See for example The Use of Colour in Language Syntax Analysis) Such presentation features can safely be ignored in assessing the results of the trials with CODIL and MicroCODIL. In addition counting, arithmetic, formal logic and the more advance mathematical techniques are cultural features alien to our hunter gather ancestors – and to all animals – and are not be considered as as part of the basic memode model – although the ability to acquire such skills is relevant to any model of the evolution of human intelligence.

Extending the Network

The CODIL network model was not designed with the memode approach in mind and there are several ways in which it needs simplifying to make it more like the the memode model of the brain. The first is related to what could be considered the “bottom level” when actual information is stored. This has been discussed in From Neural Nets to Language which also discusses some of the ideas in this post in more detail. The CODIL Item “MURDERER = MACBETH” effectively indicates that MACBETH is a character string which is a member of the set MURDERER. For the memode model this becomes a linked pair of memodes “MURDERER -> MACBETH” where MACBETH is now also considered a set, which may have members. This can then be extended to provide a broader model than CODIL normally supplies:

It also emphasises the need to do further research to consider the specialist areas of the brain, such as the the parts that process sight and sounds, as a continuation of the decision process, and a consideration of the boundary between conscious and unconscious thought. There are also some interesting aspects of the move between serial and parallel processing and the problem of self referential problems.

CODIL as a programming language

CODIL is a bottom up system which simply scans lists of lists of lists of items and makes simply deductions, with the advantage that there is no need for an explicit pre-defined global algorithm for the task. This means that there is no formal distinction between items which can be used as “data”, as “logic tests” or as “commands” - because any and every item (with a few special cases concerned with input and output) can validly be used in any of these roles. Exactly the same applies to the memode model.

But what happens if the person who wants to use CODIL has a task which can be pre-defined? This is perfectly acceptable in that the user can provide lists of items which he only wants to be used in a particular way – in effect as a program. In the simplest case this “program” could just be a single item to be interpreted as an instruction to start by comparing one list with another – and the effect can be very far reaching because of the recursive nature of the decision making process.

There are limitations to how far this approach can be taken because CODIL uses associative addressing (unlike conventional computer systems which use numbers for addressing purposes) and as such cannot easily handle algorithms which are based on arrays, although a very limited facility has been built in for handling items with multiple values. While most CODIL “program” are very small the language was used to write one very significant “program” - called TANTALIZE.

TANTALIZE was designed to solve logic problems and was named because it succeeded in solving 15 consecutive “Tantalizer” problems (now renamed Enigma) as they were published weekly in the New Scientist during the mid1970s. It was also used to solve many of the similar problems that appears in the Artificial Intelligence literature at the time. Using the terminology of the time TANTALIZE was a series of production rules, represented in CODIL, which queried a human user about the logic problem to be solved. The result was to generate a new series of production rules which were optimised using a version of CODIL with a learning facility. The optimised version of the rules were then “obeyed” to produce the correct logical solution in something approaching the minimum number of steps.

The paper of TANTALIZE reported on how the approach compared with other Artificial Intelligence research being carried out at the same time:

Compared with other problem solvers, the TANTALIZE problem descriptions are almost always simpler, the time to process problems is often more than an order of magnitude less, and with a mere thousand [lines of code] TANTALIZE provides an unrivalled range of heuristic search and other facilities. Its comparative weakness in algebraic manipulation is more than compensated for by the ease with which it can be extended by the use of CODIL to carry out a wide range of operations.

There is no way that the complex set of rules such as those that made up TANTALIZE could evolved on a memode network dependent on dynamic trial and error learning – but the fact that the architecture can support complex algorithms is important when we consider the evolution of human intelligence. In particular there is the real possibility of generating English language statements from the "MACBETH" network illustrated above 

7. Evolving more Brain Power

The basic memode model assumes that animals which can modify their behaviour based on experience must have a brain that can (a) recognise and remember objects in the real world, (b) link together objects which are associated and (c) can use this information to make simple deductions. If we then assume that the human brain is little more than a scaled up version of the animal brain we need to ask what changes need to evolve in order to support the the kind of intellectual activities humans are capable.

In increasing the “capacity” of the brain one could increase the number of memodes, the number (and perhaps the nature) of the links between them, the number of memodes (the Current Context) that can be active at one time, and factors relating to learning and forgetting – some of which relates to the depth of “recursion” which can be supported (i.e. how complex a single chain of thought can be before it starts to forget).

The important thing to remember is that, in evolutionary terms, there is no point in having a brain which has more capacity than is needed, and one of the biggest costs will be that associated with learning. The bigger the brain the longer the childhood learning period, the bigger the investment the parents have to put into rearing and “teaching” the youngster, and the more hard-earned knowledge that is lost when the brain dies. This suggests that if one can reduce the cost of learning and/or the ability to transfer significant amounts of knowledge from one generation to another, the costs of having a larger brain fall while the benefits increase. If this happened the bigger brain could make it easier to learn and transfer information and the one has what one might consider an auto-catalysed reaction, where the result of the brain getting bigger is to make it even easier to even bigger.

So can we identify a possible trigger point where developing a bigger brain became more economic, triggering an auto-catalysed growth. In general terms the appearance of language must be involved, and it is relevant to see how far this could have been based on the simple memode model.

In fact there seem to be two potential trigger points. The first relates to the point where the information passed from an adult to a child is more reliable than the information that the child can learn for itself. When this happens there are evolutionary advantages in the child bypassing the normal learning mechanisms of the basic memode model and directly accepting what is said as gospel. But these is already an “express learning” mechanism in any animal's brain for handling life-threatening events – for the simple reason that you cannot afford to learn by trial and error – and it is probably relevant that in human teaching there is often a “carrot and stick” element in the teaching process. Such a change in learning mechanism will significantly speed learning and reduce the instruction time and could explain the sudden change of gear in the way young children learn once they have got a fair grasp of language.

It is interesting to look at the cultural implications of a switch to "speed learning."  In evolutionary terms the quicker the exchange of information between parent and child the better. It helps the process if a child asks its parents questions but in the days well before before encyclopedias and public libraries some of the questions – such as “Why does that butterfly have eyes on its wings?” would have been unanswerable and would have no survival value. Rather than waste time or loose authority by admitting the answer is unknown there could be an advantage to the parent in having a stock answer to cover all such questions – such as “the invisible spirit in the oak tree did it.” Because of the way culture works, within a few generations a supernatural origin myth would have come into existence. In addition, because the speed learning mechanism is related to ensuring animals avoid danger, supernatural stories which promise a choice between an unbelievably attractive heaven and the horrors of eternal damnation in hell would be particularly attractive. Of course the religious may not like the fact that what they think was a “God Gene” only exists because humans evolved to accept as correct any statement (true or false) spouted by a suitably authoritative looking human, and that some individuals are more willing to absorb unverified (and unverifiable) information as being correct.

The second mechanism relates to “programming the mind.” The work with CODIL, and particularly the problem solver TANTALIZE show that something akin to the memode model can actually be used to support very complex algorithms, although it is not obvious how a individual could create such algorithms using simple learning techniques. On the other hand once language has advanced to the stage of being able to say “here is a recipe for doing ...” you may only need one person to stumble on the way to do something – and transmission of the new information processing technique becomes cultural. Areas where some “programming” help could be relevant would be the handling of numbers (expanding to much of mathematics and formal logic) and the syntax structures of the many different human languages.If this is the case there is no need for any special biological evolution to support such human mental activities.

This all suggests that the human brain could work using virtually identical decision making mechanisms to an animal's brain, and that most if not all of what seems intelligence is due to cultural factors relating to learning technigques. This does not mean that other changes are not significant. In this context the recent work on the SRGAP2 gene is interesting. This gene is connected with the connections between the neurons, the form SRGAP2a is common to chimpanzees and humans. However the gene has duplicated itself twice with the second duplication probably been relevant to the split between the Australophithecus and Homo lines. The duplication of the gene appears to have introduced a step function in the number of connections between neurons, which at the very least would have affected the cost effectiveness of the brain – and which might, for example made some differences in the nature of the links. Other changes associated with the size and capacity of the brain may have also had a step effect.

8. What Next?

Of course humans are something special in terms with what they have done on this planet, and the information they have collectively assembled. But how far is all this due to culture and how far are humans exceptional in purely evolutionary terms. Among the mammals we have giraffes with very long necks, elephants with prehensile noses, bats that fly, and legless whales which can swim deep in the sea. The purpose of this post is to argue that from the biological point of view humans are just mammals with big heads (in both senses of the phrase) and that demonstrate that a simple memode model, supported by blue sky experimental research on CODIL, plus changes in learning stimulated by language and culture, may be sufficient to explain our intelligence.

Of course I am aware that there are very many more questions to be answered, and outstanding problems to be solved, before the ideas can become ether firmly established or disproved. If I was a young scientist trying to find out how to bridge the “Black Hole” in the centre of brain research I would be raring to go. But of course I am not. I have long retired, having given up active scientific research years ago, in part for medically related reasons, and I have neither the (life)time, energy or resources to follow the ideas through to a polished conclusion. However I can still “think research” and if anyone wants to explore the ideas posted here in more detail I am very happy to exchange information on this blog, or in otherways brainstorm the issues.

Earlier Brain Storms
  1. Introduction
  2. The Black Hole in Brain Research
  3. Evolutionary Factors starting on the African Plains
  4. Requirements of a target Model
  5. Some Factors in choosing a Model
  6. CODIL and Natural Language
  7. Getting rid of those pesky numbers
  8. Was Douglas Adams right about the Dolphins
  9. The Evolution of Intelligence - From Neural Nets to Language
  10. The Limitations of the Stored Program Computer

No comments:

Post a Comment