Another approach to artificial consciousness
Abstract: The author considers some issues related to non-computational non-algorithmic cognitive structures. A new approach to investigations is presented in the article. According to the author’s opinion, this approach can lead to successful creation of an artificial brain. Philosophical ideas presented in this article were conceived while working on the experimental cognitive structure. The author also describes social consequences that may result from this kind of developments. Key words: artificial intelligence, AI, cognitive system, consciousness, consciousness system, non-computational system, artificial brain.
The main purpose of the article is to motivate researchers to begin practical work on non-computational cognitive structures. This article is not philosophical in the true sense of this word. It is the reflection of philosophical conclusions made in the course of the engineering project on cognitive system . The author (together with other investigators) has developed a computer model of the non-computational system capable of solving logical tasks. This system has successfully solved all logical tasks from the book of Raymond Smullyan «The Lady or the Tiger?» . Successful development of this system as the first stage of the project opened potential ways of creation of a much more powerful system. Its capabilities won’t be limited by logical problems only, will include a wide range of intellectual tasks, too. Currently, the author is working on the second stage of the project. The paper does not provide technical details or proofs of the correctness of the proposed approach. Only a fully operational model of such non-computational system may serve as an ultimate proof of concept. Such model however has not been yet developed to its completion. Still, our intermediate results show that achieving the design goals is not an unthinkable task.
2. Is there any progress in the field of Artificial Intelligence (AI) up to date?
Creation of Artificial Intelligence remains an attractable focus of investigations and development for more than fifty years. Almost all known works dealing with AI are based on the computational approach. Despite strong and continuous efforts, the researches have not shown essential achievements in this field yet. With lack of real achievements, the term AI has often been mentioned in the context that has little to do with an intellect itself, at least the way we understand it in humans. We can say that the term AI has become seriously compromised. Because of that, it is reasonable to call non-computational non-algorithmic systems (those resembling the essential features of the human intellect) in some other way. Such new terminology is gradually appearing. In the present paper, we’ll call such systems “Artificial consciousness or cognitive systems”. A suitable example of the intellectual activity is a programmer’s job. This choice is very convenient for several reasons. First, it is quite evident that this activity is intellectual. Second, it does not require outstanding abilities (in contrast to e.g. proving complicated theorems). Third, advanced programmers realize that their work is based on some simple principles, which however they cannot explain. Finally, in the modern world, an automation of intellectual efforts of all programmers would bring a tremendous economic benefit. As we see, the programmer’s job is intellectual, non-exceptional, not terribly complicated and very important for the society. Despite the economic importance of such coding-assisting systems, neither of the type has ever been built. This is an important indicator of the fact that there are no existing AI technologies, following the principles of the natural human intellect.
3. Reasons of AI related to designs failure
There are several important reasons of the lack of progress in the field of AI. Most important one is a misleading focus on computational principles while creating AI. Such trend is definitely related to the great success in the field of computer engineering. From materialistic point of view, the human brain is the only intellectual object or intellectual instrument, presently existing. Almost everybody agree that human brain operates as a non-computational, non-algorithmic system. Since the very idea of an intellect appears so complicated for the researches, it would seem logical for them to model just what we already have in our heads. However, vast majority of developers, for reasons of simplicity, base their AI designs on the principles of computational systems. This contrasts to creating the systems in a form of uniform non-computational, non-algorithmic structures, similar to human brain. The second important reason of failure is a modern methodology of scientific researches. Almost always, a new research is started based on one or more already existing theories. In the field of physics such an approach seems reasonable since any accepted physical theory describes phenomena existing in nature, and all theories must be verified by experiments. In the field of AI, however, all existing theories are speculative. None of them helped us to build a brain-like structure (or at least a computer program that would demonstrate the properties of the human intellect). In essence, this fact proves an improper choice of starting assumptions. Thus, is it worth taking the theories which do not reflect the nature of the intellect, as a basis of future research?
4. A chain of model tasks.
Above reasoning causes a question: what should we do in order to develop AI similar to the human intellect? The answer is to look closely into what natural human intellect is. Let us demonstrate how this can be done. Ask yourself a question – or provide a model task - that requires a certain degree of intellect to be answered. Then, try to imagine what could be done to answer this question. Most likely, the first attempt (iteration) will not result in the design decision about AI. However, this first question may lead to another, simpler question that needs to be answered first before approaching the original one. If one keeps repeating such iterations for long enough, and each subsequent iteration prove successful (a simpler question is generated correctly), one may be able to find a sequence of the simplest basic model tasks. These tasks must be simple enough to ensure the possibility of automatic solution. Then, moving back along the same chain of problems it should be possible to automatically answer all intermediate questions and finally the one that started the iterations. The author’s proposal is not to try solving complex problems by applying a certain selected theory. Instead, one should try to find a sequence of increasingly simple problems so as to secure successful solution of the simplest one in the chain. Then, by generalizing the method succeeded in the simplest case, and adopting it to more complicated tasks, move backward along the same chain, until some practically or theoretically useful ones start to be resolved. The important advantage of this approach is that regardless of the starting point (initial model task), one inevitably finds sufficiently simple subproblems which can already serve as models for real systems. A drawback of the proposed technique is related to considerable difficulty of choosing a proper chain of model tasks and high probability of errors in intermediate decisions.
5. The measure of intellectual complexity of tasks
In order to find a chain of model tasks, it is necessary to establish a measure of their intellectual complexity. To be more precise, such measure should tell us which one of the two tasks is simpler (or more complex) and how much they differ by complexity. As a criterion of complexity, we choose an amount of additional context information we need with respect to the amount of the data contained in the problem statement itself. Let us explain it. For example, if the task formulation is related to the idea of ‘time’, the system won’t succeed with the problem unless it also understands the meaning of ‘time’. In the first stage of the project the author was working on non-computational system capable of solving logical tasks. This system was based on logical operations. Logical tasks require very little additional information with respect to logic structure solving these tasks. Therefore, in the first stage of the work our model task had a low complexity compared to the complexity of the task processing structure. However, the capabilities of this structure appeared to be limited since it was only able to solve logical tasks. To develop the potential of the cognitive system, it was decided to base the subsequent designs on the idea of a ‘set’. Reasoning and motivation leading to this decision is omitted here for the sake of brevity. Author expects that this choice helps to dramatically improve the potential of the system.
6. Sets as a basis for a cognitive system
After specifying set as a basis of cognitive system, all tasks which were considered ‘low complexity’ in our first design, have advanced to those of medium complexity. Such change was caused by the fact that the new system involved more complexity even to implement simple logical operations. That is, even to repeat functionality available in the previous design, it required for us to teach the new system on what the logical operations are. However, the efforts needed to implement logical operation in set-based structure, proved useful. In the course of development, a number of methodological, nearly philosophical problems were addressed, closely related to the nature of cognitive structures. Since the idea of a set is more general (and abstract) than logical operations, it can help us to describe such categories as ‘time’, ‘space’ and others. The author expects therefore that the new system will be able to solve a wider scope of problems compared to its first version. A chain of model tasks has been chosen for implementation of the second design of the system. The simplest task of the chain was description of logical operations without using the idea of logic operations. It was successfully solved. The most complicated model tasks – in frames of the second design stage - include solving logical tasks and discussing the solution with human. Another example of complicated problems is development of some types of electronic circuits, while supporting interactive communication with human. Depending on success in this second design project, we’ll look forward to more available problems. The success in the second stage of the design would open a possibility of approach the third stage. This one would supposedly include the system that understands the idea of numbers, mathematical operations, and is able to write programming code by request, similar to a programmer who makes his coding by following technical specifications.
7. Occam's razor and orthogonality of system bases
At first glance, it is not very difficult to e.g. express logical operation through the concept of set. However, this should be done in a way that makes this representation well aligned with entire chain of subsequent model tasks. Therefore, not only the design work should be well accomplished, it should be done with elegance. Occam's razor  is the most important principle in designing the system. The author tries to reduce the number of used fundamental meanings (essences) as much as possible. The fewer essences the system contains, the more perfect it is. While working on the system, we were constantly looking for the set of maximally abstract and “orthogonal” essences which would allow us to minimize the required number of essences. By orthogonality, the author means the capability of essences to create different combinations with each other without constraints and contradictions. Such process of working on the system is very complicated. One has to repeatedly redesign the entire system. However, in case of successful findings, the system gets simpler and sometimes starts to demonstrate new unexpected features. This way, the system can be perfected endlessly. Of course, at intermediate stages the system does not look entirely perfect, as it should be at completion. Such imperfection does not prevent the system from solving model tasks at the current stage. However the lack of refinement could become an obstacle in future development.
8. The model of the human consciousness
In order to build the structure with artificial consciousness, we should have at least a rough model of natural consciousness. The model must reflect only those features which are important for building the artificial structure. Author has chosen to base the design of non-computational system, on a three-level model. The upper level corresponds to the «reasoning» level of consciousness, the middle one is about «revelation of hidden regularities», and the lowest one is an «unreasoned motivation». Reasoning level is the level of consciousness formed naturally much later than the two others. This level is a distinctive feature of human beings, but higher animals may possess it to a certain degree. Reasoning level has the ability of making logical conclusions, it is not something just based on logical operations, as we understand them in Boolean logic. This level is the main distinction between human beings and animals. In common understanding, it is the logic reasoning associated with an idea of human intellect. «Hidden regularities revelation» level is a more ancient level which is typical not only for human beings, but also for most of animals. Most likely, this level was formed during evolution of sensing organs in animals. This level is responsible for pattern recognition. It is probably based on specialized brain structures that are much older than the structure of reasoning level of consciousness. These structures should be heavily based on something close to ordered sets. For example, a two-dimensional image sensed by an eye is initially kept as unprocessed two-dimensional array of points or pixels ordered by their coordinates. After processing, the brain does not read visual information as a mere set of pixels. However, the only way the brain can convert a set of pixels into something sensible, is to find common regularities or patterns, typical for certain objects that belong to the same group or class - even though such objects may look differently. For instance, if a person is shown a number of different images of the letter «A», this person will be able to recognize the images of this letter, never seen before. Since such processing is vitally important for all living creatures, and presumably appeared long before human beings, it should be based on some kind of a built-in mechanism. Having appeared as visual and aural perception of external information, this ability over time has spread on hidden regularities revelation level. Not only human beings, but also most of animals possess this ability. Thanks to revelation of hidden regularities, animals can adapt their behavior to constantly modifying external factors, if they follow repeating patterns. Compared to animals, human beings have most developed hidden regularities revelation level. This level - apart from other functions - is responsible for revelation of regularities (as e.g. in IQ tests) and intuition (making a decision based on previous experience). Motivation level is the most ancient. It is possessed by all creatures except for primitive single cells. Let us exemplify this. There are two strategies of surviving in associations of living creatures: cooperation and antagonism. For instance, an animal may take certain resources (food, etc) from others of same kind in order to survive. In contrast, animals of other group cooperate in their efforts to get the food. The first case can be exemplified by lone predators, the second - by ants and bees. Let us treat social and antisocial behavior as one of the manifestations of «motivation» in human beings. People possess both these features in different proportions. A person may achieve prosperity in a criminal way or, by obeying the rules of society and cooperating with other people. Both ways lead to success and are absolutely logical. Thus, a choice can be made only on illogical basis. It is a person’s inborn characteristics which can influence one choice or the other. The author has been convinced in existence of such level of consciousness by the experience of working on the system implementing a set of logical functions. Such systems sometimes allow many logically consistent different solutions, and there should be a mechanism of selecting the one from many available. Another example is behavior of domestic pets. We all know that pets, like people, have different characters. They may be kind and sincere or, on the contrary, fierce and sly. This fact proves that the human beings and animals have similar mechanisms at this level of consciousness.
9. Creation of artificial structure by modeling human consciousness
The author does not insist on the absolute correctness of the proposed human consciousness model. However, this model is only necessary for choosing the tasks needed in design of an artificial consciousness. The author believes that logical reasoning itself is based on the mechanisms of revelation of hidden regularities and logical generalization. However, by author’s opinion, the work should start from developing a system capable of simulating functionality at the upper (logical) reasoning level. It is quite possible that operational model of such level will demonstrate certain limitations (won’t be smart enough). However, if the structure capable only of reasoning (but not of the revelation of hidden regularities) successfully works, then it would be feasible to start working on the structure implementing revelation of hidden regularities. In addition, there is a probability that in the process of working at the upper (reasoning) level we’ll get enough understanding of the mechanisms constituting the medium. Or inversely, the mechanisms of the upper level could be used – at least at the beginning - for the development of the structure at the medium level. After creating cognitive systems including two levels, we’d have a system that is sufficiently intellectual, but not independent: it will only respond to and execute human requests. If one includes motivation into such systems, they will become capable of setting up their own goals, not necessarily coinciding with humans’. Possible outcomes will depend on people designing such systems, in first place, on their morality.
10. Model of artificial cognitive structure
Our thinking reflects different aspects of the outside world. We can say that everyone’s brain continuously constructs an informational model. Since human brain contains nothing except neurons, such model could be represented as a neural network. Neurons and the links between them construct an informational model. From now on we will only mean an informational model at the reasoning level of consciousness. Let us imagine that instead of human brain we have an informational model in an artificial structure. If most important properties of the model of brain correspond to those of the informational model of the artificial structure, then one could say that the artificial cognitive structure is created. In 1950, Alan Turing provided a definition for Artificial Intelligence . Despite the fact that the definition is not very clear (or strict enough) it is believed to be the best. Turing said that if in the process of communication with artificial system, a human is unable to detect as to whether he is communicating with a human or artificial system, such a system may be considered an Artificial Intellect. Based on the argument already provided in this paper, the author proposes more concise formulation of this idea: «If an artificial cognitive structure, in the process of informational communication with a human and outside world, is capable of constructing and developing the informational model to the degree that allows it to answer human’s questions, appearing in the course of this communication, then such system may be considered intelligent» We should mention here some developments in the field of AI which, being similar by terminology, are still different by its content from the one presented by the author. Since there is nothing except neurons and their networks in a human brain, the terms «neural networks» and «semantic networks» can be associated with the author’s project. In the beginning of the article we have already noted that, most likely, there are no successfully completed works in the field of AI. We only note that the term «neural networks» is being used in a great number of works dealing with pattern recognition . The term «semantic networks» is also quite common in various scientific publications dealing with networks creating links between the nodes containing information . It seems true that any system resembling the brain is a semantic network, but not every semantic network resembles the brain. The first stage of the project could be considered as a non-computational design related to Boolean satisfiability problems (SAT) . Boolean satisfiability problem can be formulated as finding such values of the arguments which would make the logical expression true. Since SAT problem is NP-complete , it does not allow an analytical solution. In general, it may require a full search going over all possible combinations of arguments to reach the solution. In this work the author exploits a model in a form of a network of notions connected by links. Each node-notion possesses a number of properties. First, it reflects certain realistic notion (somewhat resembling the way the human being imagines it). Second, each element also represents a set since every notion possesses the properties of a set. All nodes and the links between them have the minimal number of characteristic properties, common to all nodes and links. Due to these properties, the system is able to develop itself so that the new nodes could be created or removed without an external algorithmic intervention (The algorithm works only because it is a computer model of a non-algorithmic structure, but not the realistic one.) The structure develops itself based on external information which becomes available for example in the process of communication with human.
11. Types of human activity modeled by the proposed system
Let us consider the types of human activity that can be well represented by the proposed artificial structure. First, this is engineering and scientific activity. If we provide a system with the description of a certain phenomenon, such as principles of operation of a particular device, this system will construct an informational model of this device and will be able to answer questions related to this phenomenon. If we describe a set of components, for example, chip configuration and a set of instructions for microprocessor, the system will be able to build the required design – electronic circuit - by considering the combination of different operation conditions. Second, it is an interaction between different people or systems in everyday life or in sport. For instance, consider a football game. A player constantly and continuously constructs an informational model of the current state of the game. Besides, player makes assumptions about probable behavior of other players. If informational models developed in the brain of players from the same team are similar, they can guess intentions and coordinate their actions to each other. Similar mechanisms control human’s behavior in public areas, such as a street, shop, or when driving a car. Third example is a comprehension of the text written in a natural language, and arty translation of this text into another language. While reading, a human’s brain constructs an informational model. This model allows it to correctly evaluate the contextual meanings of words and form the information from the given text. Based on this information, it can form a text in a different natural language. In addition, if the translator is able to catch not only direct information but also the style of underlying wording, (e.g. a style common in criminal society) it becomes possible to reproduce the text in another language that correctly represents both the original information and its style. The model is likely to work poorly in such areas as composition of music and playing chess. Music composition can require much larger involvement from the level of consciousness representing revelation of hidden regularities. An effective chess playing, in a way a chess master can do, may too become possible only with sufficiently developed subsystem responsible for the revelation of hidden regularities. At the moment, the author is not working on this level. One can teach an artificial non-algorithmic system on moving skills however at the current state of development, this cannot not result in an efficient solution.
12. Constructing of hypotheses in the process of thinking
In the process of thinking, a human not only constructs the current informational model, but also creates many hypotheses which do not contradict to it. Without such hypotheses, building of the model would be impossible, not to mention its development. For instance, when manager formulates a task to a programmer, the programmer should first understand the task, that is to build an informational model of the task. Even at this stage, a programmer’s consciousness generates many hypotheses. Most appropriate of them allow a programmer to construct an informational model. Then, in the process of working on the code the programmer creates other hypotheses (speculates upon various ways of solving the problem). Most likely, when solving the task hypotheses require more efforts and more branches in an informational model than on initial stage of comprehending the task. At the end, if the programmer succeeds in his goal, he forms an informational model according to the following principles:
- the model must correspond to the initial task (does not contradict to it)
- the model must correspond to programmer’s professional skills
- the accomplished task must be a part of an informational model the programmer constructs in his brain
- the program operates in accordance to given specification
It is sometimes believed that human brain uses a minor fraction of its resources. The author thinks that this assumption could be wrong. The thing is that in the process of thinking the brain has to build simultaneously a large number of hypotheses; in many cases this happens automatically. This process itself consumes tremendous resources, and this fact becomes likely considering characteristic features of human brain: it contains billions of neurons however physical processes are not exceptionally fast; they take up to tens of milliseconds.
13. Other speculations related to non-computational cognitive systems
Scientific projects in the field of non-computational cognitive systems can cause dramatic consequences for human society. Some statements are presented below without any proofs or explanation. This is only a small part of what can follow from successful development.
- A successful work on cognitive structures can change the basis of human lifestyle. It can lead to a dramatic economic prosperity
- Changes in humanitarian ideas may happen. Particularly, essential changes are possible in religious issues (the author does not assume a vulgar atheism)
- Even now, uniform non-computational systems could be successfully combined with existing technologies. For example, unlike modern processors it would be possible to fabricate one system on one silicon wafer (since such systems are much more tolerant to defects due to their uniform nature)
- We can assume that the informational processes in the Universe accelerate over time. The transition from lifeless world to the one full with living creatures took billions of years. The transition from living to thinking took a considerably shorter time. One may suppose that the human being is only an intermediate link in the evolution that eventually leads to creatures with much faster thinking.
- It is more likely that creatures capable of spreading through the Universe will first appear on Earth, rather than come to us from outside.
- Talking about artificial creatures, we understand that the problem of consciousness is least understandable compared to other issues associated with the development of such creatures.
- The successful work on artificial consciousness will allow us to understand the nature of human reasoning. It will show the limits and ways of better teaching the people.
- Cognitive structures will help solving the problems in basic sciences (such as physics and mathematics). Perhaps, human abilities in these fields are close to their potential limit.
- The works on such projects however pose a great danger for humankind in case they become instruments for antisocial groups or even governments
- However, the greatest danger is inability to lead this works to completion before humankind becomes extinct. The fact that human civilization is to disappear is certain. It is only a matter of time that could take hundreds, thousands or millions of years to happen. If contemporary civilization dies without an advanced successor, one may say that the existence of a human civilization on the Earth was meaningless (this is by the way the question related to the purpose of life).
- One of the problems typical for such investigations is our inability to voluntarily speed up to or activate them. The development progress in such systems is non-monotonous. It is similar to proving most difficult theorems. The success can be proved only by creating a fully operational computer model of non-computational system. It is hard to imagine the possibility of financing such development with its usual schedule of plans, reports and other paperwork. Such bureaucratic approach may only lead to low quality of intermediate solutions and compromise the development as a whole.
The work on the system based on mathematical logic (Boolean functions) showed that it is capable of solving mainly mathematical logical tasks. Successful – by author’s opinion - accomplishment of this stage opened possible ways of further development of non-computational cognitive systems, with a possibility to solve broader scope of problems, typical for human activity. At the present moment, the author is working on a system based on the notion of sets. The author came to the conclusion that for a successful work in the field of cognitive structures it is necessary to follow an approach which differs from ones used so far by the majority of scientists. First, we should develop cognitive (or AI) systems based on the ideas of non-computational approach. Second, the very approach should be changed in the methodology of works on cognitive structures. Most general principles of cognitive system operation are more of a philosophical than technical nature. It is unlikely that these principles may be worked out without an experience of practical development. However, an accurate formulation of the most important principles would allow us to develop a set of criteria of assessment, and guidelines in the process of developing cognitive systems. References
-  http://brainlogic.org/
-  Smullyan R. The Lady or the Tiger? And Other Logic Puzzles. – New York: Alfred A. Knopf, 1982.
-  http://en.wikipedia.org/wiki/Ockham%27s_razor
-  http://en.wikipedia.org/wiki/Turing_test
-  http://en.wikipedia.org/wiki/Neural_network
-  http://en.wikipedia.org/wiki/Semantic_network
-  http://en.wikipedia.org/wiki/Boolean_satisfiability_problem
-  http://en.wikipedia.org/wiki/NP-complete
-  A New Approach to Artificial Intelligence: Non-Computational AI
-  Artificial Brains