On The Nature Of Intelligence

How we think, learn and solve problems

Background

The week before my tenth birthday in 1965, I asked my eldest brother to teach me the game of Chess. During the second game that we played together, he pushed a pawn forward in preparation for what I now know as a Bishop Fiancetto. In other words, by moving his pawn out of the way he was subsequently able to move his bishop into a position attacking the centre.

I didn’t get that at the time however, so afterwards I asked how he knew to make that move, because it didn’t make sense in itself. He said that he had learned how, via experience. I then asked how we learn stuff, and it transpired that nobody knew. So I set out there and then to discover how we learn. I was so naive then that I didn’t realise that my task was thought to be very difficult – I just carried on regardless.

This was during the very early days of computers, so I barely knew anything about them. All that I had for research purposes at the time were my limited knowledge of Evolution, studying my own thoughts – which literally gave me a headache – and a book that my sister bought me as a birthday present, called “The Chess Mind” by Gerald Abrahams. It was advanced material, even for expert Chess players. I persevered however, trying to figure second-hand what the thought processes involved must be like. Fortunately the book had many annotated games, which helped.

Later on during my teens, I discovered another book, this time via the local library. It was “Chess Programming” by David Levy. Although that was useful, it set me out on a mental diversion for several years, devising low-lever heuristics or guidelines that would help with selecting candidate moves. These are required in Chess and other complex activities, because there’s a humongous number of possible board states in any game, which even with today’s super-fast computers cannot be adequately explored in a reasonable time. All of the modern chess-playing programs need such human insights to focus only on those moves that are likely to be significant.

It was only many years later – 2010, I think – that I returned to my original goal of working out just how we learn. It was around the same time that I extended my objective to understanding just how we solve problems, as that was the focus of AI research.

I held back from publishing my findings until now however, because a truly intelligent machine can be extremely dangerous; with their perfect memory, extended sensory apparatus (e.g. infra-red) and sheer processing speed, such machines can run rings around their human creators.

The more conventional researchers in the field are getting close however, and I don’t want that enormous power being held by large, unaccountable corporations or governments. So I have decided to note down my thoughts, in the hope that others will be able to make use of them or at least understand the dangers that we face.

My career has been that of a Software Engineer, so I am also able to draw on that resource during my journey of investigation. I still don’t fully understand how it all works in Nature, so I have ‘cheated’ a little and adopted some IT ideas to help out.

Key Points are highlighted like this.

Of necessity, these ideas are a little imprecise and sparse. See the Python code if you want more details.

Overview

What Is “Intelligence”?

Problem-Solving

The definition of Intelligence that I have adopted is the ability to solve problems, in any domain. According to Dr. Edward de Bono of “Lateral Thinking” fame, a problem is the difference between what one has, and wants or needs. Solving a problem is therefore a process of getting to the goal state from the current one, via various ‘moves’, or intermediate states. Being able to solve problems in principle is utterly useless however without relevant domain knowledge. For example, if you don’t yet know how to read then you will probably struggle with learning anything much.

Consider a super-being that is able to answer any question put to it, via its vast experience. It may therefore appear to be extremely intelligent, but in fact it does not actually solve any problems but instead provides solutions from its endless supply of answers.

Now contemplate a truly intelligent machine that has just been built, but which knows nothing about the real world. Even it could solve problems in principle it is unable to do so, if only because the thing doesn’t yet have the ability to understand the questions posed to it! Like a baby, in other words.

Effective Intelligence therefore requires the innate ability to not only seek solutions but also learn from experience and observation.

Basic Assumptions

Evolution/Simplicity

My resources initially were very sparse. I knew a little about Evolution, and Concepts. I assumed from the outset that Intelligence evolved from simple beginnings, as it seems that Evolution favours Simplicity initially, due the lower energy expenditure involved. I was also unable to cope with anything more complex at the time!

Memorising By Association

I have also assumed that we memorise by association – thus any and all (perceived) events that occur at the same time are linked to each other, at least to start with. As a reminder, that includes spatial relationships that are ‘scanned’ into memory.

It appears that the brain builds a mental map of its world, as perceived via the sensory organs. We then use that imperfect model to navigate and orientate ourselves in physical reality.

Basic Operation

The brain/mind is considered as a ‘blank slate’ from birth, able to perform only a few basic operations. In practice, babies know about their mother’s heartbeat and seem to recognise their parents’ voices, because they can hear whilst in the womb.

One of the most primitive operations is that of Comparison – checking whether two signals or attributes are the same or not. This simple ability plays a large part in ‘pattern processing’.

Knowledge Acquisition (Learning)

A basic life-form like an amoeba can notice events in its environment. Those could be threats or opportunities, like a potential food source for example. Such events therefore get noticed, especially if they are repeated over time. N.B.: such repetition occurs via intermediate perception of events, which themselves may be generated by the body. Just moving ones eyes around to look at various things will generate a succession of ‘signals’ for the brain to process for example.

Memories of such events tend to fade away over time, and they need to be repeated if they are to be recorded over the longer term. I guess that any retention needs to be justified as it were, due to the energy and storage costs involved.

In the domain of Chess, those events obviously include the actual moves made during a game. At a lower level however, they also include noticing – and learning – the board features of squares, diagonals, ranks/files etc. The Move Number is itself also grist to the mill, and the initial board state at the start of a game is obviously a frequently-repeated observation. A lot of active learning is therefore required when initially learning the game, via asking suitable questions. That in turn implies a prior and more passive absorption of the native language, which in my case is English. It’s all about events, repetition and memory in essence however, so studying a single domain like Chess can also yield insights as to how we learn in other contexts.

Concepts

What Is A Concept?

A concept may be considered as ‘blank slate’, which can have various attributes assigned to it, and links to other concepts. Such characteristics can include basic physical aspects such as a spoken word, a colour, simple shapes etc. Further, a relatively simple concept can be extended to ‘include’ other concepts, such as eyes in a face.

Concept Formation

A concept is formed via repetition of events, initially those that are perceived via the senses. A new/unknown, event as perceived is recorded as a basic ‘proto-concept’, along with a Value attribute and a unique timestamp/identifier. The latter also has the benefit of recording the creation date/time, perhaps with millisecond resolution, which is useful for IT debugging and reporting purposes.

Each subsequent event of the same kind is mapped via its characteristics onto the corresponding concept, thereby reinforcing it and making it stronger, more persistent in memory.

As with their foundation-stone neurons in the brain, each memorised concept will decay in strength over time, at least in the context of Short-Term Memory, or STM. During a Learning phase, their activation energy is maintained by repeated triggering of new events. Thus we have learning by rote, still practised today in many infants’ schools I believe.

Anything can be taken to excess however, and the brain has a way of dealing with over-stimulation: it just switches off. Why waste time and energy on yet another car alarm? So the Concept triggering is probably ignored for ‘endless’ repetition of events.

Compound Concepts

The key mechanism that makes the architectural model viable.

Each concept has a unique id., and so it can therefore appear as an internal event in its own right, when activated. This means that higher-level, more complex concepts will evolve.

In Chess for example, a ‘Castling alert’ could be triggered via a combination of conditions – the King and Rook unmoved in their original places, with a ‘chunk’ of two or three squares in between. After learning from many recorded games, the corresponding ‘mental map’ concepts would be formed. A suitable software program could then highlight such a feature, and prompt for a name to be assigned to it – ‘Castling Conditions’ perhaps. Likewise, features such as Diagonals, Rank and File could also be named, building a foundation for subsequent natural language learning in due course.

For a simple example, consider learning a sequence of characters, in groups of three: ABC … ABD … ABE … ABX. The AB pairing is repeated throughout, and once the signal strength has exceeded a predetermined threshold, the AB group is recorded as a proto-concept and has a unique timestamp assigned to it. In practice, that means that each element of the group will be labelled with that timestamp, so that all the elements are bound together by it. Once that concept creation process is complete, the AB pairs in the original input are replaced by a reference to the newly-formed AB concept.

This very important process is what enables a complex hierarchy of concepts to be built, as a mental map of the perceiver’s world.

Feature Abstraction

Imagine walking down a road, observing a single house as you walk past it. Each ‘snapshot’ taken would be unique, due to the viewing angles involved. I doubt that we record everything that we see, touch or hear, so kind of abstract model must exist in our mental maps, which matches well enough with the perceived scene to permit accurate and reliable recognition.

This may be explained when one considers the lower-level aspects involved. A window for example will typically consist of four straight lines, forming corners. Those will rarely be right-angles in the actual perceived image, but the essence of a window is the relationships between those angles and lines.

At the fundamental level of perception, we may well have straight lines and possibly circles etc. hard-wired into our brains via evolution. A Scientific American articles that I read once referred to an experiment conducted on a cat, by inserting electrodes into its brain and then presenting the poor animal with pictures of straight line in various orientations. It was noted that several clusters of neurons would ‘fire’ in response, matching each orientation as presented.

The Difference That Makes A Difference

With apologies to Gregory Bateson.

Again, in the Chess domain, assume the Square concept is already known – already on the mental map – but the concepts of Diagonal and Rank, File are knowledge still to be acquired.

Now consider the act of visually scanning the board. When each square is viewed, it will trigger the corresponding Square concept, which in turn will energise or activate all of the attributes of that concept – colour/name and rank, file location. As usual, the learning of the algebraic a..h and 1..8 coordinate features is also done via association.

The input to the ‘Chess Mind’ therefore consists a series of Square events, with little or no delay between each event. I assume that there is an innate ‘difference’ comparator mechanism, which itself generates a ‘same’ signal with each Square event. That output is also associated with all the other events occurring at the same time. Additional inputs may also include the viewing angle!

Thus for any rank (horizontal line of squares) on the board, the scanning process will generate a series of events like Square, (left/right) Edge, 0 delay (distance), Square etc. Further assuming an ‘activation/association period’ of perhaps 1/10 sec., a ‘snapshot’ is taken of all the active concepts at the time and the whole new proto-concept is created with its own unique timestamp.

Thus via scanning of the board, a sequence of ‘match’ events is triggered, which are then reassembled in a ‘virtual space’, all of the relevant activated concepts broadcast their status to all other existing concepts. If none match the current arrangement of events, then a new concept is formed.

The process of comparison can also be applied to the timestamps themselves, yielding results of ‘before’, ‘same’ or ‘later’. This is required to permit codification of temporal sequences, such as a musical ditty. A similar pattern also occurs with the above spatial example, so the eventual Rank concept may consist of Left Board Edge, ‘a’, Square 1, ‘before/after’, Square 2, ‘before/after’ .. Square 8, Right Board Edge. I have omitted the interval delays for brevity. The ‘before/after’ elements are derived from timestamp references, which translate to spatial relationships on the board itself – assuming a left-to-right scanning process perhaps.

Language Learning

A key aspect of this system is that each concept can not only be uniquely identified via its timestamp, it can also be named or labelled, via a simple attribute assignment. Thus a Cat concept has a (textual) name and other attributes, such as Colour, number of legs, various physical perceptions like sounds and behaviours. A major one will of course be the spoken word ‘cat’, which enables the concept to be activated whenever the ‘cat’ utterance is perceived.

Now consider the sentence “The cat sat on the mat”, which may describe a realistic scene of course. Each individual word is assumed to be already known to the perceiver, but not yet the sentence. As described above, each ‘word event’ occurs in the given sequence, so the built-in Comparator process generates delay values, and event counts may also be valid at this stage. As all of the words are held in STM and are active, they are all associated with each other, together with the system-generated components. So the first word ‘The’ will be linked to the last word ‘mat’, along with all the others.

The important point to note here is that the mind model has predictive power, so if you hear “The cat sat ..” then your brain may well expect those words to be followed by “… on the mat”. Further, the key words ‘cat’, ‘sat’, ‘mat’ are all linked to each other, corresponding in essence to the physical reality. The common (frequently-repeated) words like ‘the’ have little definite meanings (strong associations/links) in themselves, but simply help to form the expected sentence structure.

In due course, as more words are added to the ‘Matrix’ knowledge base, it becomes possible to model any real-world environment in any language, and convey that information to other perceivers, if they know the same language and context. The Compound Concept mechanism permits rich, meaningful maps of the world to be developed. A face ‘contains’ eyes, a nose, ears and so forth.

Problem-Solving and Planning

As above, a problem/goal/objective is defined in terms of a starting state and a desired state. Both will have their own collections of activated concepts, which correspond to candidate moves in the case of Chess. Thus it may be possible to work forwards towards a solution or backwards from it, though the latter is often more useful as it closer to the objective.

A major aspect of the whole system is Values – in IT terms, numeric values (and associated behaviour) that are imbued early in the teaching stages. Common examples may be Hunger, or the need to replenish energy supplies. In Chess of course the ultimate objective for each player is win the game, or more accurately not lose it. That in turn translates in paying attention to the King (both sides), or passed pawns. The King in particular will have neighbouring squares associated with it, simply because the piece can only move one square at a time.

Thus, as soon as possible, the players will be contemplating moves or positions that will limit the opponent’s options, by blocking the escape routes for example. For the top players, all of this thinking material – concepts – is obtained via years of intense study, stuffing new arrangements or even entire games into their mental maps, along with their own good/bad evaluations.

The actual cogitation process is much like that of a conventional Chess program – a process of Trial and Error, basically – but using learned concepts rather than hard-wired moves etc. Thus the end goal of Checkmate will activate Check or other relevant move concepts, based on the pieces left in play (concepts active in memory) and a ‘playbook’ of previously-acquired knowledge. Those candidate moves will then be juggled around in STM, encountering various weak proposal along the way, then backtracking to a better state and then repeating the steps, with other material in mind. Eventually the highest-value available move will be found, and then played on the physical board.

Thus the process of problem-solving consists of clearly defining the desired goal first, then contemplating various manoeuvres that may yield a promising approach, based on experience. All quite simple! The obvious complexity of cogitation is due to the sheer volume of data items and the relationships between them.

Conclusion

A Truly Intelligent Machine (TIM) will be able to run rings around us, due to its sheer storage capacity and processing speed. We have already witnessed the rise of ‘clever computers’, which far surpass human abilities in certain fields of endeavour. There is a lot of funding provided by military groups, and it seems that there is at least one company offering highly-targetted, cheap Killer drones – aimed at individuals perhaps via face recognition. I estimate that the IQs of such machines will one day reach the high hundreds, or even thousands. By comparison, Mensa members are around the 250 mark.

As usual, those who pay the piper call the tune, so any AI developments will inevitably be biased towards concentrating yet more money and power in the hands of small number of people, who are largely unaccountable to to Society as a whole.

I hope therefore that people in general will find this information to be useful, and even attempt to build their own AI counter-measures! I hope to publish some Python code soon, but I have currently lost access to it 🙁

© 2023 Richard Prosser - Licence: Creative Commons Attribution 4.0 International (CC BY 4.0)

Leave a comment

Your email address will not be published. Required fields are marked *