Plan to develop artificial Cogitation – what has to be done

2012-12-13

Plan to develop artificial Cogitation – what has to be done

The long-term goal of artificial intelligence science is to develop an artificial intelligent mind. That means, an intelligent system which is able to learn new facts and consequences, and to perform independent cogitation which leads to an ideas of solution of its goals.

Before we can succesfuly develop such system, we have to solve several problems.

1) Covering AI theory

We will have to create general artifical intelligence theory / algorithm / model, which would fit for the purposes of intelligent system. Such model will have to satisfy these needs (among others):

  • Self-organization:

    System must be able to organize it's inner structure to some extent. This extent will be dependent on what functions will be hard-coded in the model's structure and which functions the system will learn.

  • Must be able to store all kind information in the same way.

    An intelligent system will need to treat all it's conscious knowledge the same way. A „piece of knowledge“ is a subject of infering with other knowledge, and the way of infering itself is also subject of further infering.

    Talking about infering suggest the way of representing the knowledge – by associating different terms. A term (idea, conception, notion) can be any kind of information about any object at any level of abstraction: „the sky is blue“ is represented by associating a term „sky“ with a term „blue“. These terms are created by regular observing the environment with receptors. The association „the sky is blue“ is created by observing the environment, which activates the term for „blue“ and the term for „sky“ (can be perceived as „that area that is above everything else“).

    The model also has to store information about dynamic processes: About its own actions, observated actions of other objects, and the effects of these actions. Again, this will be stored as association between the action and its effect, using the terms involved in action.

  • Processes represented the same way as facts.

    To be able to learn beavior, the system must be able to store processes the same way as facts and to handle these processes the same way. This is implicitly said by the previous point.

    //Update:// A. V. Gavrilov:

    The basis of reasoning lies in operating with fuzzy images by means of a process of associative recall of images (see Principle No. 1). At the end of the process, a choice of certain operations is carried out (recalling of it): it is therefore possible to associate the successful choice (the solved task) with the focusing of attention, the start of operation as programs of operation motor neurons, etc. Thus the selected operation is involved as a tag in the further process of reasoning.

  • Must be able to infer from the stored information.

    Cogitation will be done by infering all types of information from all parts of the system:

    • Long-term memory – facts learned, generalized information about classes of objects, stereotypes, long-term goals.
    • Mid-term memory – recent actions and their effect; „where am I“, mid-term goals („what's my today's plan“)
    • Short-term memory – current „context“ of the system.

    These will not be separated in some systemic way; the separation will be rather fuzzy and will ensue from the usage of the particular information (or association). This will follow the well-known principle of association re-inforcement.

    The cogitation would probably happen mostly in the short-term memory part.

  • Must be „serializable“ to some extent to be evolvable.

    It's almost clear that an artificial cogitating mind will not be designed. In my opinion, humans are not able to design as complex structure as human brain. The way to design it is to design a general principles of how the structure works, and then let it evolve, using something like genetic algorithms.

    During evolution, a general structure will „crystalize“ and some of it's parts will be hard-coded in the genetic information; other parts will be described vaguely and left as a subject of self-organization.

    To be able to store the structure and properties of the structure, it has to provide means to store its every aspect as a serialized genetic information.

Of course, there are more things that such AI theory should satisfy.

Now let's create a simple example of how a primaeval man can think in this model. It shows how the information could be stored and how inference could be done.

Based on some experience, a primaeval man could create this association:

action „crush“ subject („stone“ property „sharp“) → effect „cuts“, effect „divides“, effect „hurts“, …

(Note that the meta-words like „property“ and „action“ have to be represented somehow in the system, too. Using them here this way is just for brevity.)

The terms at right side could be associated depending on the situation in which the action was performed – thus the association itself would be associated with terms which were activated in short-term memory in that moment.

Now the man sees other man hunting an animal with a spear. He watches and creates an associations:

property „sharp“ ↔ effect „pierces“, effect „hurts“
object „spear“ (abstraction 0.8) <= ⇒ „sharp“
object „spear“ (abstraction 0.8) ↔ „wooden stick“ (abstraction 0.9)
object „spear“ (abstraction 0.8) ↔ goal „hunted“ subject „animal“ (abstraction 1.0)
action „stab“ subject „spear“ <= ⇒ goal „hunted“ subject „animal“ (abstraction 1.0)
(and many others…)

Then, some day, he desices to go hunting. „hunt an animal“ term is activated and associated as mid-term goal. (TODO: how?) Using the association created before, the „excitation“ reaches the „spear“ term. He gets an idea of „making a spear“. Using associations

object „spear“ (abstraction 0.8) <= ⇒ „sharp“
object „spear“ (abstraction 0.8) ↔ („wooden stick“ property „long“, property, abstraction 0.9)
action „crush“ subject („stone“ property „sharp“) → effect „cuts“, effect „divides“, effect „hurts“, …

he infers an idea of creating a spear by crusing a wooden stick with a sharp stone. Having no better ideas, he associates this idea as a short-term goal, having all the terms and associations used to get to idea activated. As he goes through the environment, using his receptors he recognites a stick that activates the „long“ term. That (re-)activates the short-term goal structure of associations, so he stops, gets the stick crushes it with a sharp stone. So he makes it sharp, efectively making a spear. So he reinforces all association in the structure used to infer the plan to make a spear.

Then he succesfully uses the spear to hunt an animal, and having done his mid-term goal, the association between a goal of hunted animal and the structure of making the spear is reinforced. And because having a hunted animal is very important goal, the primaeval man is very happy this reinforcement is very strong.

This example demonstrated how the intelligent cogitation could work. (Maybe I should not suggest my own solutions for the cogitation system in a list of its necessary properties, but I am not good at creating structured documents ;-)

//Update:// General description of cogitating intelligent system's principles is in A. V. Gavrilov's The Principles of Action of Intelligent Systems. This description overlaps my ideas, but could hardly put it in such easy-to-read, consistent document.

Estimated time: 10 years to a complex AI theory tested in simulated (virtual) environment.

2) Computing power

Based on currently accepted estimates, a human brain capacity is approximately 10 000 TeraFLOPS. Although, this estimation is very rough and can not be closer in principle – human brain does not perform floating point operations.

Today's fastest centralized supercomputers, like Ranger, have about 504 TeraFLOPS.

Today's fastest distributed computation structures, like SETI@home or BOINC, have computing power over 1 200 TeraFLOPS.

To imitate the human brain behavior, first we will have to get to a level when __common__ computers will have brain's computing power. Why?

As described above, the structure in which the cogitating system will work, has to evolve – it's very improbable that humans will design it. Thus, we need a computing power at least for:

  • running __all__ individual cogitating systems that will take part in evolution
  • computing their „fitness function“, that is, evaluating their „intelligence“
  • performing genetic operations on their genes.

Seeing the sum of computing power needed, it's almost clear that to develop, evolve and teach artificial cogitating system, we will need much more computing power than human brain has.

Estimated time: 45 years to first potentially cogitating, but not learned mind situated in simulated environment; another 2 to 5 years to teach it and see whether it's learning at all.

However, the second point of the list above is discutable: Will we really simulate system's envi­ronment, or will we place the system into the real environment with sufficient number of receptors and effectors? That depends on what actually the artificial mind will be for. Should it live in the internet and do some job in its virtual environment? Or should it act like an educated monkey in real-world? In both cases, simulation might be either insufficient in one case, or it would need even more computing power than the system itself. Thus, it's more likely that we will learn the systems in the real world.

That makes a good question: How will we evolve the system and place it in the real world in the same time? The answer is „pluggable brain“. The physical „body“ of the system will consist of it's receptors and effectors, and the „CNS“ will be unplugged / downloaded from it for genetic operations, and plugged / uploaded into it after genetic operations.

This way, we can skip the time needed to form the body. However, remember, that we do not perform genetic operations on the system's knowledge, but it's structure. So we can't skip the time needed to teach the system. Depending on the growing complexity of the system (not the task), its learning will be longer. „Not the task“, because our goal is not to create a system that will solve some „task“; our goal is to create independent cogitating system. We will let the system face very various tasks, from simplest to more and more difficult. The teaching of the system will be similar to education of a dog, or to bringing up a baby; it just might be more intensive and faster.

Corrected estimated time: 30 years to first prototype of potentially cogitating, but not learned mind situated in real world; another 5 to 10 years to teach it and see whether it's learning at all.

The last point of the above list will doubtless be the least computing power demanding:

Current estimates of number of human protein-coding genes is 20,000–25,000. I suppose we will not be able to be as efficient, and the length of coded data will rise exponentially with the size of the precisely described parts of the system (cogitation center, visual recognition, sound processing), minus the sizes of parts that will be described vaguely (e.g. memories, „fact“ storage).

However, it is likely that genetic operations will not be done using a computing device of any kind. Instead, biological and chemical processes will be used; most likely, it will be done using natural ways:

Let's say there are more ways to encode information about how to form a system. Imagine a world (e.g. Earth at the time when RNA principles were forming, at the level of self-reproducing „crystals“). In this world, all ways of information encoding exist, materialized and working, at some time. They spread, forming some systems. At some point, the systems, and that means their way of inforation encoding too, will necessarily start fighting for resources. Which systems have better chance to survive? The ones whose principles of „genetic information“ are more efficient: Duplicates faster, and in case of mutation, this mutation is likely to introduce a new feature of the system instead of corrupting whole information.

So, with some level of certainty, we can say that if there are more ways of encoding information about system, history choosed the best one – RNA, and later, DNA.

So, we can include another need – „fast“ reading and writing of genes and their processing.

3) Genes reading / writing and their processing

As „deduced“ above, RNA / DNA is probably the most effective way of storing an information about a system. If we adopt it for our system, we will have to develop the means of fast and reliable reading / writing of DNA.

Fast means that the process of reading / writing will not involve high-tech laboratory, many rounds of segmented scanning, time-demanding concatenation and uncertain parsing.

We will also need the means of genetic operations of mutation and crossing.

  • Mutation will probably happen accidentaly, but we may also rise it's level using e.g. focused radiation.
  • Crossing is likely to be performed in specialized biological cells.

Note that we will probably not use DNA as instructions for cells which proteins to create; DNA will be just our „storage medium“.

Estimated time: 20 years.

4) Principles of computation

Current computers still operate on the basis of binary logical circuits and work on the principles of the Turing Machine.

However, Maass and Markram have recently argued that

„in contrast to Turing machines, generic computations by neural circuits are not digital, and are not carried out on static inputs, but rather on functions of time“.

We have two possibilities: Come up with mental model that could be run using binary computers, or design new kind of computers which would be more appropriate for this task.

Since binary computers are said to reach their limits soon, the second option is more probable.

Estimated time: 15 years.

0