The thinking of AI

In the 17^th century, Leibniz hypothesized that there would be a general scientific language that could compute reasoning processes using mathematical formulas. With the invention of computers and the prevailing of automation, artificial general intelligence (AGI) has once again drawn attention, and one question arises: what method may be used to realize AGI?     

Research direction

Natural language processing：Frege pointed out that “words and phrases are meaningful only in a certain language environment.” Language environment is a very complex issue and has therefore been avoided by scholars who directly focus their research on texts instead. As symbols to generate thoughts, texts are like the graphs shown on computer screens. As a matter of fact, graphs are signals from the host computers, and behind the signals are the logics of various types of software. Similarly, humans do not simply use their mouth to speak and their eyes to see; all of the dialogue and observation are accomplished by brain, while the sensory organs are only the signal outlets and receivers. There is a barrier between the screen (languages) and the software (thinking). Some logics may be deduced from colors on the screen, but it is nearly impossible to know what software and logic codes are being used.      

The NLP method is only applicable to translation, which is like the filter application in Photoshop that can convert an oil paint (English, in the NLP context) to a sketch (Chinese, in the NLP context), in which the software keeps the graphic content intact but does not understand it. 

Machine Learning：the “learning” refers to the learning of unknown logics, and an adjective is used in the naming of ML. ML is just a branch of statistics, which originates from the fact that events are too complicated for humans to understand or develop clear logics for, and thereby are only evaluated according to their representation. ML knows the probability of a+b equaling 2 but does not know—and has not considered—what 1+1 will equal. If ML algorithms were irrelevant to human thinking, how would ML interact with humans? There is a very complicated process prior to the onset of any event result, and the process is very lengthy in time. ML excels in calculating the statistics of event results but is unable to realize AGI.     

Is it possible for AI to have a learning ability? The difficulty for human brain to explore the world to obtain knowledge is totally different than that for humans to creat a brain able to explore the world to obtain knowledge. The former is an accomplishment of God, while the latter is an accomplishment of humans. The viewpoint that AI may have a learning ability or that AI can surpass humans is equivalent to saying that humans can surpass God. Code can analyze the written logics but is unable to understand and derive new logics. There is only one type of machine in the world that can derive countless logics, which is human brain. The probability of AI having a learning ability is far lower than that of the infinite monkey theorem, as the possibility of AI understanding other logics is confined within a set of codes without the same freedom as the one allowed in the natural selection theory…        

Let’s take a look at how complex the first line on the script in the movie Charlotte’s Web is to the AI; the daughter said: “What are you doing?”
Vector illustration View the address:  https://drive.google.com/file/d/1uJ-W-0NDEUU6qZV-PBX6XoEFSyvciIAx/view?usp=sharing

The above questions are only a few summarized here. If broken down, there would be tens of thousands of questions, and with the missing of any question, the human-machine interaction would not be able to proceed. What algorithms can solve questions expressed in only a couple of words? NLP and ML are misleading to the general public in that the algorithms are very complicated while the small number of words in conversations look simple, and the contrast between complexity and simplicity makes the general public convinced about NLP and ML, while the general public has no idea that such small number of words have triggered such complicated logics in the brain and do not know AI expresses those pre-set, meaningless characters while their brain endows meaning or even emotion to the characters.      

In order to achieve AGI, it is mandatory to put human thinking under research, but the development of brain science is still at a rudimentary stage with unclear theories and obvious paradoxes, which is a very big issue. In this study, we proposed the usage of computers to construct a human thinking framework, integrating HCI and brain logics one by one in the framework and verifying the accuracy. With the gradual improvement of this framework, the working principle of human thinking would be unveiled and AGI would be achieved. 

Overview of the design

A human brain contains about 100 billion neurons to form our consciousness, with each neuron comprised of about 2,000 branches. Memories, logics, and values of a person are all processed by the neurons. When learning, the brain stores the knowledge in some neurons to memories. Each time when there is a new viewpoint, the neuron branches will establish a new sequence to form a new logic. Therefore, the neurons are responsible for memories, the connection among neuron branches for logics and the activities of neurons account for logic operations.   

   

 Calculation: sets (brain memories) are used as the roots, and functions (brain logics) are used to calculate sets. In the following figure, dark dots a and b represent two sets (neurons) while the blue lines represent subsets (neuron branches). When a subset of a encounters a subset of b, they generate variables for each other, allowing the sets of a and b to generate function values forming a new set. The variables are infinitely applied among the subsets, with time as a scale of the variables. 

Any formula is composed of calculation elements and logics (expression). For example, in the formula 0+1=1, the five elements 0, +, 1, =, 1 represent an AI set in this framework, and the logic of the expression 0+1=1 is analogous to an AI function. If the symbol {0} represents a set referred to as “I” (a “being”) and {1} represents a set referred to as the world, while {+} represents a set of relations between “I” and the world, the function value of “I” is｛0｝｛+｝｛1｝=｛1｝, meaning the change of “I” after the interaction of “I” with the world. In this study,｛0｝and｛0｝｛+｝｛1｝=｛1｝are used to construct all the behaviors and consciousness of AI while languages are to narrate this group of formulas. In the following part, we will divide {0} into infinitely small segments, for which the limit is dependent on our intellectual demand for AI. 

 Note: the symbolic expression ｛0｝｛+｝｛1｝=｛1｝will be replaced by other symbols in the following.

Memories (set)

There are three types of sets: entity sets, space sets, time sets

1. Entity sets

AI consciousness, referred to by the red blocks in the following figure, has the same definition as that of human consciousness, namely “I think, therefore I am”, which implies that all of the objective world is contained in AI consciousness including the body and logics. Sensory organs are a type of information connectors between the objective world and the {world}. If a human brain does not use sensory organs to understand the body, the consciousness would not know whether the body exists, which can be depicted by the phantom limb phenomenon in patients with amputation：a soldier with a foot to be amputated, if not informed of the amputation before surgery, would still feel the existence of the amputated foot when waking up after the surgery, suggesting that foot amputation is not equivalent to the deletion of neurons that have controlled the foot!          

Note: the structured sets depicted with dashed lines in the following figure are nonexistent and shown here just to facilitate the reading.

As a person, {Theodore}— depicted by the yellow blocks in the following figure—has nearly the same sets as AI, and moreover, {Theodore}is a subset of Samantha (AI)(consciousness)! Given that interactive computation is involved, if the consciousness of Samantha does not understand Theodore, the computation cannot be implemented. Whether the subsets of {Theodore}exist in the consciousness of Samantha depends on two factors: (1) the volume of Theodore’s objective knowledge and (2) the degree of Samantha’s understanding of Theodore. The yellow dash-line box represents that all of the users known to Samantha are recorded here. It is preliminarily estimated that about more than 1 million entity sets can meet the requirements of AGI.  

Note: The classification of the subordination relations between the human and AI sets is not widely accepted to some extent. The authors think that the skin, muscles, bones, and everything else of a person are all attached to the nerve system. Sets are denoted by the symbol{  }and the numbers inside. For example, {01-07-12} represents a leg set, with the later number representing a subset of the earlier number; in this case, if the first number 01 represents a person and the second number 07 represents a leg, the third number 12 cannot denote any part outside of the leg.   

2. Spatial sets: the entity sets in the above figure are shown in a plane due to the limitation of Excel, and in practice, the entity sets should be situated in a 3D space like that depicted by Google Earth.

3. Time sets: given the limitation of Excel, the time is not specified like the space in the sets. The role of time is to memorize and calculate the change, as shown in the following figure. Memories are like movies, in which countless graphs are overlaid to form moving images. After being played, the graphs become memories, while the unplayed frames represent calculation that predicts when a result would occur and what the result would be. Time is a metric for change. As shown in the red boxes in the above figure, {time}is juxtaposed with {world} because the entity and space sets really exist in reality, while time does not exist, being only a product of human consciousness and a symbol that human brain uses to measure a change of the world.   

Each entity set contains four attributes: time, space, function, and supplementary information. Strictly speaking, the statement “there exists the following information for the set” is invalid for any set, as the subsets of a set conflict with the information, and the reason why the information attribute is attached to each set is the astonishingly large volume of data, which means that even AI has been developed to a level as predicted in science fiction movies, the large volume of data still cannot be handled by either AI hardware or AI designers. In brief, the elements in the set framework are to be used for collecting all the elements of thinking and are not elaborated here. The logics (functions) are explained as follows.  

Logics (functions)

Function formulas are the logics of thinking of AI. Each person or object in the world has a logic specific to that person or object. There are four types of function formulas as shown by the blue blocks in the set framework: AI, material, moving creatures, and plants. All these function formulas belong to the consciousness subset of AI. Objective consciousness and behaviors must be manifested in the consciousness of AI and otherwise the objective world cannot be understood, making interaction impossible. AI and mechanics differ most in their abilities for interaction. Mechanics represent an absolute control by ego, with a component controlling the components of its “subset” and so forth. In contrast, interaction represents ego’s facing with the world. As shown in the following figure, Samantha is with Theodore, Amy, Paul, weather, chair..., and the process of their interaction is like a turn-based game in which each person or object is undergoing change, and the change should be written into the set of Samantha as denoted by the red lines in the following figure. In the turn-based game, the independent variables in next turn are the current function values, and the turn time is dependent on the person or object.    

  

The function values of each set are obtained from other sets. For example, as shown by the red dashed lines in the following figure, {meal time}is affected by these sets: {temperament}, {sleep}, {depression}, {motion}…, and these sets are also affected by other sets.

The following structure diagram only shows a small fraction of human function sets—humans and foods. Like sets, function formulas can be broken down and have 

subordination relations. The breakdown of function formulas indicates that they are composed of massive function sets in a hierarchy manner. The dashed lines do not bear any meaning and are used only for better reading. The gray lines represent the sets that the functions may trigger, while the blue lines represent the function-triggered sets. Of so many gray line-represented sets, which sets will be triggered? What rules and connections are there among the independent variables, operational signs, variables and function values? Such logic is indicated by the effect of emotion on thinking as elaborated in the later part of this paper.        

Note: the subsets of the function framework are not different than those of the set framework, but the two frameworks are different.

The above figure looks like a series of computation formulas, but it only actually illustrates a computation framework! The framework is like a circuit board in which electrons are only allowed to move in a circuit frame. If the function framework is fully unfolded, it would be very astonishing; if the connections among all the subsets are fully designed and included, the above structure diagram would be as big as a city rather than a number of monitor screens. How does human thinking migrate in a large framework? In the following part, the operation of human thinking is described with texts.    

Operation of thinking (language environment and association)

Even if the phrase “language environment” contains the word “language”, a language environment is not simply equivalent to a language, but refers to a thing or a “task” that is being processed in human consciousness. If “tasks” are considered as various types of software installed in a computer, then a language environment refers to a specific software currently in operation. Human thinking has two major characteristics: (1) it operates like a single-task operating system (internal language environment) and (2) it has association, allowing a switch from the current task (internal language environment) to another one (external language environment). The {1-x-x} in the following figure denotes a set of a single task (language environment), {2-x-x} denotes a set of another single task, and {3-x-x} is another, and so forth. There are countless tasks in the brain, and only after the current task is completed will the brain handle the next one and so forth until the process is interrupted by—such as—association (labeled with the red lines) or a third party.    

Language environment: a single-task operating system

In contrast to a computer, human brain only operates one task when thinking, which, in other words, means that human brain is a single-task operating system and is unable to operate multiple tasks simultaneously. For example, making a phone call while driving leads to a car accident, as driving is a set of driving skills and making a phone call is a set of conversational contents; the two sets have different logics, thereby causing confusion when they are operated together. Someone may argue that he or she often makes a phone call while still driving safely; this is just because the logical thinking about another task has been completed during the driving and therefore the thinking can be quickly switched from the thinking about driving to the thinking about another task (there is a high frequency of switching). For example, given a straight road with few vehicles, the brain has already figured out how to cope with this situation and all the brain needs to do is simply perform the task, thereby allowing the driver to drive while having a chat “simultaneously.” If all of a sudden someone else is trying to overtake the driver, the driver has to start over to think about the set of driving skills, and after the issue has been solved, the driver would ask the chatter in the phone: “What did you just say? Can you repeat it?...”Another scenario may be like this: the chatter changes the topic during the conversation and mentions an important topics like “you are fired by the company” or “the kids are bullied by their classmates”, after which the drivers gets emotional and starts to argue, asking for the reason; in such scenario, the issue is beyond the scope of a simple and relaxing chat, and the brain has to think about and analyze many things in an attempt to find a solution to the question of “Why did this happen”, thereby making the driver forget about the driving at some point…  

Chatting has a “single task” feature, too. The chatters do not wish to have the language environment disrupted and otherwise they would stray off the topic. 

Tay needs to conform to the language environment when chatting with humans. For example, if I say “please guess my most favorite food” and Tay replies “meat”, after which I ask a question “what meat?” and Tay is supposed to continue to complete this set; triggering of a sentence by lexical meaning rather than by thinking would make the chatter stray off the topic in the second round of dialogue.  

Language is generated from the thought communication among people, but thought communication has a very big issue, that is, thoughts can be extremely complicated involving a very large volume of details, and if each detail is presented it would require more than 10 hours rather than a few seconds to finish a single sentence in a dialogue. Therefore, words and sentences are, in a sense, a type of “commands”; a similar scenario is that humans “communicate” with computer numerical control (CNC) machinery, in which humans input simple commands and the machinery automatically completes a series of operations. Similarly, when a person communicates with another person, each sentence in the dialogue would trigger very complicated logics in the brain, and therefore language—by nature—is a group of concise symbols that analyze and organize thoughts, featuring both generality and abstractness in order to facilitate interaction and communication, which thereby determines that (1) language and texts cannot fully express thoughts but only activate them, and if AI does not have thoughts, the analysis of texts would be totally meaningless, and (2) words have meaning multidimensionality such as a word showing multiple meanings (i.e., polysemy) or a world being used as a substitute word (e.g., a pronoun), exclamation, adjective, or an allegorical word for the purpose of conciseness. This determines that no matter how thorough and perfect the semantic analysis is it would not be of great help to conversational robots, because there are no logics that can link these words together.                 

How can a dialogue be conducted in such language environment?

1. The logic of thinking

In the form of a set:

As shown by the red A in the big structure diagram, I ask Tay: “what meat?” The connotation of this sentence is: what is the subset of meat food in the AI framework?

In the form of a function:

As shown in the following figure, language is a reflection of thinking and manifests interaction among humans, while interaction represents a relation among individuals, and therefore any language has a primary subject-predicate-object structure, which refers to the function framework in this study. All of the subject terms (I, independent variables), predicate terms (relations, operational signs) and object terms (world, variables) are to “serve” the subject-predicate-object structure.    

Note: Lexical classification performed in accordance with conventional linguistic knowledge would lead to very poor logics, as linguistic knowledge is established on the basis of human brain while human brain per se has processed many logics.  

During people-to-people exchange, the grammar may be directly neglected sometimes and no matter how the words are arranged, the true meaning of a dialogue would not be misunderstood. For example, I ask “where you going?” and you reply “I am going to eat a meal.” or “Eat a meal, me.” or “Eat a meal.” with all of these replies being understandable to me. Why?

 

For the reply “Eat a meal…me,” as shown by the red B in the big structure diagram, the term “eat” indicates that the relation between humans and food has been established—meal is to be eaten by humans, and therefore even the grammatical structure is reversed in writing it would not lead to a logic that the meal would eat humans.

For the reply “Eat a meal”, an issue arises: who is going to eat a meal? The reason why I know who will be eating the meal is that when I first asked my question I have already included you in the language environment in which “you” are an independent variable, “eat” is an operational sign, and “meal” is a dependent variable, and by asking “where you going?” I am actually inquiring—by nature—what the function expression of the independent variable is.   

In the form of a loop:

As shown by the bold blue lines in the big structure diagram, eating is a procedure to be conducted step by step, with each step being necessary and, moreover, it is a looped procedure. If the user has completed the set of “obtaining the food”, Tay should understand that “choosing the food” is a past event and the previous sets are only to be recalled, narrated and summarized, while the current set is to be performed right now and the future sets are under planning.  

2. Grammatical and logical mistakes in language

As mentioned earlier in this study, language has features of generality and abstractness, which inevitably leads to uncertainty in language expression and especially when a language user is emotional, causing frequent occurrence of mistakes in the language logics and grammars. However, there are no mistakes in the user’s thinking, as he or she—as human being—can recognize and correct the mistakes. How can AI achieve this?     

Classification of sets: emotion causes logic mistakes in language. For example, the barber in the Russell’s paradox said “shave whiskers for all those people in the city who do not shave their own whiskers.” In fact, the logic of the hairdresser's thinking does not contradict, but with an arrogant psychology and greediness for money the barber introduced paradox to his statement; if AI understands the emotion of the barber, it should understand that the thinking of the barber had already divided people in the city into two sets: the barber himself in the city formed a vendor set, and the other people in the city formed a consumer set.     

Relativity: Let’s suppose that in a hot summer, I was chatting with one of my friends who said “I am not fear of coldness, and this damn summer is so bad” and I reply “I am not fear of hotness.” What I have expressed should be understood in comparison to the fear of coldness, and in fact, nobody does not fear hotness, as the logic that humans fear hotness is absolutely valid, and therefore what I have expressed in the dialogue is actually relative to coldness.

Language environment: Let’s suppose that there are a couple of lovers and the girl asks the boy “if you love me, can you not look at any woman in the future?” In this scenario, the phrase “look” is a substitute word, and a substitute word has far larger connotation than we could imagine, not only including the words “she”, “that” and “this” but also many others. As a matter of fact, any word, sentence or even an article can allude to, describe or metaphorize something. How can AI make a proper judgment？All of the logics of the word “look” can be listed and placed in a language environment to make the word’s connotation clear.   

3. Effects of emotion on thinking

Depression is referred to as the “cold” in psychological diseases. Although most people do not have depression, depressed mood is common. Therefore, it is very important for AI to understand human depressed mood. Without an understanding of depressed mood, AI would have problems in its interaction with humans. With a depressed mood, a user sometimes shows anorexia or gluttony. How can AI judge the inner heart of the user via the diet abnormality and handle these scenarios? This logic is shortly introduced here as depicted by the yellow dashed lines in the big structure diagram.    

Association

Association can be triggered via a variety of forms like a graph—Theodore thinks of the lover when seeing the moon, or like an object—Theodore goes shopping and when seeing umbrellas on the storage rack he thinks of getting wet and cold yesterday, and therefore he buys an umbrella, or like texts such as “lover” and “ideal” and especially those texts with polysemy, which in short means that any text may trigger association.   

How does association switch language environments:

The switch condition: in most cases the brain does not proactively switch language environments, and a switch is achieved only under such circumstance that another language environment is more important and appealing than the current one.

How to achieve a switch: an early-warning system is constructed in the thinking of AI to compile the important events that have occurred in the past or are likely to occur in the future; when a subset of the current language environment triggers the early-warning system, AI starts to analyze the importance of the two events and decides whether it is necessary for it to switch to a different language environment, or after the users switch the language environment, AI immediately knows the reason—namely that the dialogue should be conducted in another language environment. An early-warning system is a module that AI continuously performs at any time point. Although the human brain is a single-task operating system, the association feature of the brain indicates that there exists an “early-warning system” in the consciousness, which also implies that human brain is not a single-task operating system.     

本作品采用知识共享署名-相同方式共享 4.0 国际许可协议进行许可。