How the Human Brain works – concept cells, memodes and CODIL

BRAIN 

STORMS 

Number

13

I am really Excited.

I can't wait to tell you about it. 

See Later Follow Up Discussion Papers 

From the Neuron to Human Intelligence: Part 1: The “Ideal Brain” Model Part 2: Evolution and Learning (to appear shortly)

I am currently preparing a talk How Evolution has made us the way we are – and I had a problem. You can't really assess how evolution has affected the development of the human mind unless you have a clear model of how the brain processes and stores information. I have already, in earlier brain storms on this blog pointed to the black hole in brain research where very significant amounts of work is being done round the edge of the problem but where there is no good model of how electrical pulses in the brain are translated into human language or behaviour. I have also discussed at length how work on a highly unconventional language called CODIL (COntext Dependent Information Language) may throw some light on the issue and recently introduced the term memode (memory node - see The Evolution of Intelligence - From Neural Net to Natural Langauge) to try and demonstrate a possible mechanism. However there was still a major gap in the model if it was to serve as an adequate model for understanding how the evolutionary pressures worked.

So What has Happened???

The trigger to filling the gap came from the article “Brain Cells for Grandmother” in this month's Scientific American, by Rodrigo Quian Quiroga, Itzhak Fried and Christof Koch. (see also Concept Cells: the building blocks of declarative memory functions by Rodrigo Quain Quiroga). This research involved monitoring the activity of single neurons in the medial temporal lobe of epileptic patients undergoing assessment prior to surgery. It was observed that a single cell might respond to different pictures of a single individual, while remaining inactive when pictures of other people were shown. In one case a neuron was found which responded to three different pictures of Luke Skywalker, his name (either in writing or spoken) and interestingly to a picture of Yoda, another character in the film Star Wars.

One of the problems I was facing in earlier brain storms was the relationship between the information in the working memory (called the Facts in CODIL) and the main memory which contained the items of information and the links between them. On reading "Brain Cells for Grandmother" the relationship became clear. Information is stored in the links, and the main memory and the working memory are one and the same – the difference is that the working memory is defined by the links which are currently active. I can now tie my model into the neural network in the brain and the change of viewpoint provides throws light on the following aspects of the working and evolution of the human mind..

• The “virtual” information represented when the top neuron in a memode is active is simply the sum of all the subsidiary memodes (recursively) which are linked to the memode. Thus each concept is at the top of a tree of subsidiary concepts. There is no one location where information about a concept is stored.
• The senses trigger bottom up activity in the memodes, the objects  we “see,” “hear,” or “smell” being the highest level memodes activated.
• The CODIL Decision Making algorithm maps onto a process in which active memodes can trigger top-down activity in related but initially non-active memodes. (Basically the brain can decide that if “A” and “B” are active “C” should be active.)
• Consciousness (the information we are actively aware of – and equivalent to the Facts in CODIL) is the sum of the information represented at the time by the top active memodes.
• Learning involves the linking of the top active memodes to generate a new higher level memode. (In fact it looks as if deciding what not to learn and what to forget becomes the critical factor than learning in building memories.)
• If we assume that some decisions are made sequentially - with a series of memodes triggered in a predefined order – the model could support at least a simple spoken (i.e. sequential) language. The experience with CODIL being able to handle significant non-trivial information processing tasks is relevant here. This points to an evolutionary tipping point leading to the explosive growth of both language and other special skills.

For more details read on ....

Redefining the Memode Model

The starting point of the model is that all neurons are equivalent, in that they can receive messages from other neurons and transmit messages to yet other neurons.

• EVERY neuron represents a “concept” but in some cases the concept may be at a very low level – for instance that cells in the eye has detected movement – or at a high level – such as recognising a fellow human, or even something as complex as  “An Understanding of the Evolution of the Human Mind”.
• A “memode” is the tree-like network of lower level neurons which CAN feed messages to a single neuron at the top of the memode.
• Memory is the sum total of the concepts represented by the subordinate neurons, each of which is at the head of a lower level memodes.

So basically the memode model is recursive and the complexity of concepts that can be handled depends on whether there are any limits of the depth of nesting of concepts. In addition concepts are defined by the links between the neurons, but not all links are equivalent.

A Simple Model

Figure 1

Figure 1 shows how a memode, in this case for a cat, can consist of lower level memodes relating to many different senses – and if  you are conscious of “cat” you are normally not conscious of thinking  “four legs”, long “tail” or “paws”. If someone speaks the word cat you are not conscious of the phonemes involved and if you read the word “Cat” you think cat – and not the letters “C”, “A”, and “T”. This suggests that your conscious thoughts consist of the highest level memodes that are active at that moment, and that lower level memodes are usually masked by active higher level memodes.. However to some extent information can flow down as well as up in a memode. For instance if you see a cat and concentrate on the concept you may well be able to imaging the smell of a tom cat. If you stroke a cat you may well imaging it purring. And if you want to communicate what you are thinking you may well activate concepts linked to the “Word for Cat” concept.

It is important to realise that some of the lower memodes – such as “Four Legs” or the letter “A” are concepts in their own right and will appear in many other high level memodes.

Actively Processing Information

Figure 2

It is important to realise that the memode structure must not only store information but also actively process it. Figure 2 shows what happens if a human see a cat and decides to write down what he see. The eyes start off activity which ends up activating the memode “CAT” and the text branch of the memode is activated to generate the writing of the word cat. In the same way the stroking a cat will activate the “Feel of a Cat” and this could trigger the branch “Sound of a cat” and the brain will imaging it can hear the sound of a cat purring. However a mouse might consider the question “I can smell cat and I can hear cat - so can I see cat?” and by downwards activating the “Sight of the Cat” branch the eyes are commanded to look around to see it there is a cat in view.  

Working Memory

In this model the working memory can be considered as a temporary higher level memode which links the top active memodes and the maximum number of active memodes is the maximum number of concepts that the brain can simultaneously manage. New memodes can simply be generated by making the links “permanent”. For instance if someone is looking at a picture of a unicorn Memode A might be a memode linked to the concept for “something like a horse”, Memode B might be linked to the concept for a “long straight horn” and Memode C might be linked to the written word “Unicorn”. Remembering involves a neuron developing links with the three active concepts – so that in future if any become active the links still exist. Internally the new memode does not have a name but we might like to call it the Unicorn memode.

Not all links are equivalent

It is important to realise that links are not permanent and not all carry the same weight. For instance “Like a Horse” does not automatically define a unicorn but the word "Unicorn" does. In addition learning activities will “reinforce” some links while others may prove wrong, or a never used, and forgotten. The model needs to cater with simple logic relationships such as “If A and B then C”, and also to avoid going into infinite loops. While the way information is stored is different different many of these problems have been solved for CODIL and should be transferable; some need further investigation to see if CODIL is relevant, and some need further work. Rather that make this post even longer I will attempt to cover these points in future Brain Storms.

Programming the Mind

Apart from the fact that the above examples use words to describe the meaning of memodes, and there are some references to the spoken word, all aspects of the model are almost certainly applicable to animal minds. However there is one area relating to linkage which is very important with relation to language and other skills which are unique to human.

The brain is a network of neurons and works in a parallel fashion. Some activities in the real world require things to happen sequentially – and natural language is one. Many human skills, such as making a flint axe, or writing an essay on the human brain also require the ability to order actions. It is obvious that we can do this to a significant extend, and some animals at least have some such ability. This ability is important as the ability to activate a series of memodes in order gives the brain the ability to store and run “programs”.

The experience with CODIL is very relevant here. CODIL was used to support a historical data base with biographies of 6000 people and a smaller data base of medical data on patients at a local hospital. It was able to support various interactive teaching packages, and to solve a large number of artificial intelligence tasks. In addition a great variety of small scale tests on different types of task were carried out. The basic processes are so simple that it was possible produce a logically powerful version, that worked in an early home computer with only 32K memory for everything - including the screen display. MicroCODIL, was designed for teaching artificial intelligence and other information processing ideas and was trial marketed for school and attracted rave reviews. A consideration of these tasks suggests that the proposed memode model could well be capable of processing significant quantities of information in sophisticated ways, including language.

The problem is that learning a complex sequential task means developing a lot of memodes AND getting them in the right order. This is a significantly more difficult task that remembering the association between a picture and a word for a unicorn, and I think it very unlikely that human mental activities could get as far as they have without something significant changes in the way we learn special skills and pass them from one generation to another. This suggests a key tripping point in human evolution. Once language has developed to a point where it can pass on "the program" for a task new skills can be passed from generation - being improved as it goes. As language is itself a task, once the tipping point is reached  language skills will also be passed from generation to generation. And as language improves it becomes possible to communicate more complex tasks. The result  is an explosion in both the complexity of language and other skills in what a chemist would call an auto-catalyzed reaction. 

See Later Follow Up Discussion Papers

From the Neuron to Human Intelligence: Part 1: The “Ideal Brain” Model 

Part 2: Evolution and Learning (to appear shortly)

OK - So this has been a very brief introduction to a model that is trying to provided a mechanism to explain how electrical activity in the neural network that is the brain can be linked to mechanisms for processing of complex concepts involving large amounts of information. I would be the first to agree that there are limitations and areas that I have not explored, and potentially a mountain of further detailed study which I cannot possibly do myself.

The key question is how new are the ideas. I have been searching the online literature and there is an overwhelming amount of literature which looks in very considerable depth on a single topic, in many different relevant disciplines. Many of these researchers admit there are problems in understanding what is happening in the brain outside their specialist areas. However I can't find any similar models which can link the activities of individual neurons to the explosion of language and other skills that has occurred in human evolution. If, because of my limited resources I have missed something please let me know.

As mentioned above, I am preparing a talk next week which will include some information on how evolutionary pressures probably took a mammal brain and stretched it into something working at the human level without any significant change in the basic information processing mechanism. I will post the lecture here in about a week, and an expanded version of the evolutionary tale in March. I will also be preparing further, more detailed notes, in areas which touch on matters raised by people who comment on this blog, or where more detail seems appropriate - such as the way the CODIL processing mechanisms are reinterpreted in the memode model.