The Intelligence Machine

Copyright © 2001 by Tienzen (Jeh-Tween) Gong

  1. I. Dive into the un-computable universe
  2. II. The mapping principle
  3. III. Engineering a science
  4. IV. Designing parameters for an intelligence machine
  5. V. Topological neuron
  6. VI. Two design criteria
  7. VII. Window signal and signal memory
  8. VIII. Signal registration
  9. IX. Second registration & recall memory
  10. X. Topological map memory and recall memory
  11. XI. A thinking system (I)
  12. XII. A thinking system (II)

I. Dive into the un-computable universe

With a new Epistemology, there should be a new paradigm for the 21th Century.

Please visit:
With this new epistemology, we are able to investigate the spaceless, the timeless and the un-computable universe, and one of the such a subject is intelligence.
Intelligence is the source for understanding physics, but traditional physics is unable to understand how intelligence arises.
II. The mapping principle

For human, to understand a thing (its structure, internal logic, and what not) must map it into human brain. That is, a "subset" of human brain structure must be "isomorphic to" that thing.

If the entirety of the universe (its structure, logic, laws etc.) can be mapped into human brain(s), then the universe can be understood in its entirety by mankind.

If the entirety of God (His structure, logic, essence, laws, etc.) cannot be mapped into human brain(s), then God is uncomprehensible by mankind.

This "Mapping Principle" can be a tool to investigate two issues.
We can "design" this artificial intelligent brain without knowing anything about the human brain physiology.

Can we design a brain better than a human brain?
Can this be a method to understood the structure of human brain?
Let's find out these together.

III. Engineering a science

Is engineering a science? If it is, it is quite different from the science of physics. For engineering, a problem can be resolved without truly knowing its cause. For example, a machine is vibrating very badly. Engineers can use some vibrating isolators and shock absorbers to solve the problem without knowing the cause of or eliminating the source of the vibration.

In Fictitious Universe Epistemology, we are actually "engineering" physics, and this provides many more degrees of freedom in the research of physics. "The problem" of the 20th century (the unification between gravity and other forces) was easily solved with "engineered physics", and it only needs to fabricate three parts.
Even if the real universe were absolutely different from the above "engineered" universe, this FU shows that there are other possible universes and other possible physics.

IV. Designing parameters for an intelligence machine

The moving space-time can often obscure or even erase its previous signatures. Thus, things can develop and become quite distinct from their origins. The research of an intelligent machine (brain) might not be able to discover what intelligence itself is. And this type of research is confined by its research object. We will have a much freer hand by designing an intelligent machine from the ground up.

Now, we shall first set some designing parameters or list some issues. If they don't work out, we change them.
  1. Is the known intelligent machine (brain) a source of intelligence?
    Or, is it (brain) a result (product) of an intelligence (process)?
    Can intelligence be created by or composed of something not intelligent? Human brain consists of three things:
    Which one creates intelligence?

    It would be the easiest if all three carry intelligence. Do we truly have a free hand to design intelligence in such a way? We could find this out together. This engineering process has a similar problem to the transcendence and immanence paradox of theology.
    Does an intelligence exist apart from each intelligence machine?
    Does every and each part of an intelligence machine carry intelligence?
    We will discuss these in detail.

  2. Human brain seemingly has an infinite memory.

  3. The death of a small number of neurons in the course of human life naturally seemingly will not cause a memory lose.

  4. Intelligence consists of, but not limited to, the followings:
We will not discuss the above as right or wrong issues. We will simply "design" some mechanisms or processes to perform them.

V. Topological neuron

We can design an intelligence machine by first designing a topological neuron (the t-neuron). This t-neuron has the following properties:
  1. it has about 1,000 dendrites (short connections) and many axons (long connections) to make contact with other t-neurons.
  2. This t-neuron is a topological torus (a ball with two holes). One hole let things go into this t-neuron; one let things go out from it. The hole can be as one point or as the entire surface of this t-neuron.
  3. This t-neuron has an ideal "balanced" state.
  4. This t-neuron can get into an excited state by receiving something (being fired or charged upon) through its entrance hole (the end of one or many dendrites or axons). This state can be called as waking state.
  5. When a t-neuron is fired repeatedly, it will finally reach a fatigue state, and it no longer can be fired.
  6. An excited t-neuron can discharge something (firing) through its exit hole and gets back to its balanced state.
  7. A fatigued t-neuron must go through a "reset process" by discharge something (firing others) in order to regain the ability to get into the waking state.

When a t-neuron is imbedded in a t-neuron mass, it must follow a social rule. When its neighbors are waking, it most likely be stay awake. A fatigued t-neuron cannot easily be reset when others are waking. That is, a mass of t-neuron must "reset" together about the same time.

A resetting t-neuron must, in fact, charge its neighbors. However, a group (mass) t-neuron can still be reset with the following tactics:
The bigger this t-neuron mass is, the longer the resetting time it needs.

I do not know why animals and human need sleep. But I do know that a t-neuron mass must be reset (t-neuron sleep), and there are many t-neuronal activities during this t-neuron sleep (the t-neuron dreams).

When a cell has only a handful neighbors (like a muscle cell), it can be somewhat as itself. When a cell is connected with thousands neighbors, it must obey the social rules:
It forms, at least but not limited to, three state:
It takes two processes to mediate the three states above.
Because every t-neuron is connected with thousands neighbors, it cannot function the two processes alone but must synchronize with others. This makes two states for a t-neuron mass:
Yes, a t-neuron mass must sleep and will dream.

After a memory system is designed for this t-neuron mass, it can perform the followings:
This t-neuron dream will also provide a similar mechanism for Freudian dreams.

VI. Two design criteria

I do not know whether human brain can store infinite bits of information or not. The intelligence machine we are designing should be capable of doing so. Mathematicians would disagree the using of the term of "infinite" here. I will make a definition as follow:

"R is the number bits information that our intelligent machine can store. For a (any) nature number N, R can always larger than N."

Strictly speaking, especially in terms of mathematics, R is called "finite but unbounded," not infinite. However, in my paper, I will make the terms of "unbound" and of "infinite" to be inter-changeable. I will prove that the R of my machine is "unbounded," and will simply call it "infinite." This might not be exactly correct in terms of mathematics, but it is convenient to our discussion, and it will not distort any of the issues.

My t-neuron mass (the intelligent machine) has a finite number of t-neurons (about 10 billions) and has a finite number of connections (about 10 trillions).
If the information is stored in individual t-neuron or its circuits, we can always find a N larger than R. Perhaps, this R is large enough for human's life time, but the R of my machine must be infinite.
Because my t-neuron is a life, it could die. If information is stored in individual t-neuron itself or its circuits, its death can cause to lose a vital bit of information or to disable an important memory circuit.

For the two reasons above, my design will not choose the above methods regardless of whether human brain uses them or not.
In order to resolve the t-neuron death problem above, my machine uses a group (for example, one million) t-neurons to store one "unit" of information. Thus, the death of 10 or 20% of this group will not lose the stored information.
In order to find an R to be an infinite, the finite number of t-neurons must be used "over and over" for different bits of information.
Information can be written on a "magic slate" (a toy) "over and over." Of course, it does not have memory, and I will resolve this problem later.

A magic slate has two states:
  1. as a white sheet
  2. as a written page.
It must go back to the white sheet state before a new page can be written.

VII. Window signal and signal memory

Now, my t-neuron machine has a few windows; vision window (eyes), sound window (ears), pressure window, etc..
When a chair is detected from a vision window:
  1. It generates a window signal.
  2. A group (about one million) t-neuron got excited when they receive this window signal.
  3. This excited group forms a topological map (topo-chair) in this t-neuron mass.
  4. When this window signal disappears, the topo-map (topo-chair) disappears, and the t-neuron mass goes back to its "random state" (the white sheet state).
  5. When a desk is detected, another group got excited (written state) to form a topo-desk.
  6. Topo-chair and topo-desk can share some t-neurons (10%, 20%, or...)
  7. When that same chair appears again, a new topo-chair forms, and this new topo-chair is about 95% identical to the old topo-chair in terms of their t-neuron members.
  8. After repeating the chair window signal, a "signal memory" comes alive as the topo-chair's members become consistent. Of course, this signal memory is far from a recall memory.

With this "signal memory" mechanism, two problems are resolved.
  1. The R of this machine can be as an infinite because the finite number of t-neurons can be used over and over. The information is not stored in individual t-neuron itself but is stored as a window signal topo-map (ws-topo-map). A finite number of t-neurons can create infinite number of ws-topo-maps as the t-neuron mass has two states:
    When a t-neuron mass lose its ability to return to its random state, some particular images (topo-maps) will stay and appear over and again. This t-neuron mass will, then, suffer a schizophrenia-like symptom. Drugs which can relax t-neurons back to their random state will ease this symptom.

  2. Because the information is stored in a group (about one million) t-neurons in the form as a topo-map, the death of 10% or more of that group will not lose the stored information as the topo-map should not be distorted significantly.

VIII. Signal registration

With a repeated window signal, a ws-topo-map will soon reach a fatigue state which must begin to discharge. Because this group is connected with one another via their dendrites (short connections), it is much easier to discharge via axons (long connections) under this circumstance. That is, a new topo-map at a distant place is generated by this fatiguing ws-topo-map. I call this action as registration of the ws-topo-map and this new topo-map as register (reg-map) or syntax of this ws-topo-map.

Now, a chair can generate a window signal.
If we can design a mechanism to activate the reg-map internally, this reg-map could activate the ws-topo-map without the window signal. That is, we would have designed a recall memory.

With a recall memory, we can, then, design a "thinking machine."

IX. Second registration & recall memory

Only with a "recall memory" system at hand, we can, then, design a thinking machine. Although the window signal can be registered in the t-neuron mass, a mechanism is needed to activate its reg-map internally (without the window signal) for a true recall memory system.

So far, the signal memory has a uni-direction only:
window signal --> ws-topo-map --> reg-map (syntax)
Now, four window signals are mapped into (not onto) a t-neuron mass as followings:
  1. chair --> ws-topo-chair --> reg-chair
  2. table --> ws-topo-table --> reg-table
  3. baby --> ws-topo-baby --> reg-baby
  4. crying --> ws-topo-crying --> reg-crying
These four reg-maps can be viewed as a new topo-map. When this new topo-map is getting fatigue, it generates a 2nd order reg-map (a relation-map). A relation-map can be called as reg2nd-map.

The next day, four window signals are mapped into this t-neuron mass as followings:
  1. chair --> ws-topo-chair --> reg-chair
  2. table --> ws-topo-table --> reg-table
  3. baby --> ws-topo-baby --> reg-baby
  4. laughing --> ws-topo-laughing -->reg-laughing

Obviously, the topo-map of reg2nd (chair, table, baby, crying) has a significant overlap with the topo-map of reg2nd (chair, table, baby, laughing).

The more syntaxes the two reg2nd maps share, the less the difference between the two maps. When the difference between the two reg2nd maps reduces to a level, they can switch from one to another and vice verse. I call this very-alike switching (or va-switching).

That is, after a reg2nd (chair, table, baby, laughing) is activated by some window signals, it has a chance to switch to reg2nd (chair, table, baby, crying), and it can activate reg-crying without the window signal of crying.

With the 2nd (or higher) order of registration system and a va-switching mechanism, a recall memory system is designed.

Now, we need to summarize what we have designed so far.
  1. a t-neuron, as a ball with two holes; that is, it can be charged and be discharged.
  2. a t-neuron mass (about 10 billion t-neurons), its member has 1,000 dendrites and many axons.
  3. this t-neuron mass can be excited by window signals.
  4. t-neuron will reach a fatigue state after being excited repeatedly.
  5. there are two ways to reduce the t-neuron fatigue:
  6. the t-neuron mass is a white sheet
  7. this white sheet can be written repeatedly, as it always goes back to a white sheet state when the external signal disappears.
  8. Although this t-neuron mass is a white sheet, a "signal memory" can exist.
  9. the window signal can be registered (burnt in) as reg-map at another part of this t-neuron mass
  10. The 2nd order registration forms a topo-map network
  11. With a va-switching mechanism and a topo-map network, a recall memory system can be designed.

This topo-map network is not a hardware circuit which could be found in this t-neuron mass because the background state of this t-neuron mass is a white sheet (the randomness).

This topo-map network is formed with the followings:
If we feed two "identical" t-neuron mass (having identical neuronal connections) with different window signals, their topo-map networks will be completely different. They will be two entirely different intelligent machines.

For a window signal-chair, only when the reg2md-chair has "burnt in," it will, then, become a part of the long term memory from which the reg-chair can be recalled. That is, the t-neuron mass is, now, having three different functioning sections:
  1. A section for window signal mapping.
  2. A section for short term memory -- 1st order of registration, reg-map.
  3. A section for long term memory -- 2nd order of registration, reg2nd-map.

X. Topological map memory and recall memory

We have briefly outlined the design of a "recall memory" system which starts out from a "signal memory." As we are not researching the human brain memory system, there is no issue about right or wrong on this signal memory. It is simply an issue of whether it is a good or a bad design. So, we are not obligated to provide an explanation of why this signal memory should work. However, I do want to provide an explanation on it here because that mechanism is very important for our future design works.

The forming of ws-topo-chair is quite arbitrary when the t-neuron mass receives a window signal-chair the first time. The shapes of ws-topo-chair for two different t-neuron mass could be quite different.

With a repeating window signal, two things happen to this ws-topo-map.
  1. the member of this map begins to join the group; that is, it begins to recognize the other group members.
  2. the group as a whole moves toward to a fatigue state which causes a registration, the generation of a reg-map.

Every t-neuron has over 1,000 dendrites, and we can number them as follow:
(D1, D2, ..., D1000, ...)
A t-neuron (good boy) can join:
Thus, an individual t-neuron can be a member of many different topo-groups. And this t-neuron will reduce the "activation resistance" for a "group signal" of which it is a member. If a signal in the form of (D1, D3, D8) is not recognized as a member signal, the activation resistance will be higher, and this t-neuron will not get excited until that signal is high enough to draft this t-neuron to be its member.

This membership mechanism gives rise to a "topo-map memory" which applies to all topo-maps (ws-topo-map, reg-map, reg2nd-map, etc.). Again, this topo-map memory cannot be found as any hardware circuit in the t-neuron mass.

We, now, can give a more detailed outline of our t-neuron mass memory system design as follow:
  1. The information is stored in topo-map ( a group t-neurons, numbers from tens thousand to a few millions), not in any individual t-neuron.
  2. An individual t-neuron can be a member of many different topo-maps. Its ability to recognize those different topo-maps gives rise to a "topo-map memory."
  3. With a fatiguing mechanism, a topo-map can be registered with a reg-map.
  4. With a 2nd order registration, the reg2nd-map forms a topo-map network.
  5. With a very-alike switching mechanism, a recall memory system can be designed.

XI. A thinking system (I)

We arbitrarily designed that the group size for the ws-topo-map to be about one million t-neurons. The group size for reg-map does not need that big; one-tenth of one million should be fine. The group size for reg2nd-map can be even smaller.

The smaller the group size a topo-map has, the easier it can be activated.

Our t-neuron mass does not have a power supply plug. It is "on" all the time. It is "on" even during the t-neuron sleep.

Although the background state for the t-neuron mass is a white sheet (the randomness state), an reg2nd-map can be activated even during the t-neuron sleep because of two reasons:
  1. the activation resistance for a topo-map is much smaller than the resistance of an arbitrary group.
  2. the reg2nd-group is relatively small.
So, a reg2nd-group can be activated internally, and randomly. This is an internal random activation mechanism.

The more often a topo-map is activated, the lower its activation resistance will be.

Now, there are two ways to recall a topo-map:
  1. by window signal via a very-alike switching mechanism.
  2. by an internal power activation mechanism of the t-neuron mass.
Soon, a small group of reg2nd-maps will have a much lower activation resistance compared to the others. This small group of reg2nd-maps will, then, take over the control of the neural activities of this t-neuron mass. When some topo-maps can be activated internally in a non-random fashion, a thinking machine has come alive in this t-neuron mass.

so, what is "thinking" in terms of the traditional definition? With our new intelligent machine, we can define "thinking" as follow:
  1. Although window signal can be a part of thinking, thinking is neural activities which do not rely on any window signal and is activated internally in the t-neuron mass.
  2. Thinking is not a random activation process. It is leaded by a small group of reg2nd-maps.

Thus, thinking must have a starting point (the initial condition) and a pathway provided by those initial topo-maps.

With a "thinking" system designed, we, now, are able to discuss the issues of:
  1. rational reasoning
  2. irrational reasoning
XII. A thinking system (II)

When a few reg2nd-maps light up at the same time, they form a new topo-map, and it can also be registered. For the designing concern, we can map this 3rd order registration onto the reg2nd section, but this might not be a good design. I will create a new section -- reg3rd-map and let all higher order registrations (4th, 5th, ...) map onto this reg3rd section. That is, there is no registration section higher than reg3rd. This is not a right or wrong issue but is a design choice.

Now, let's review our design again.
  1. Information (ws-topo-chair, reg-chair, reg2nd [chair, desk,..], etc..) is stored in a group of t-neurons, not in any individual t-neuron.
  2. Individual t-neuron has the ability to memorize all its membership ID. 3) An individual t-neuron can be a member of many different groups (topo-maps), such as a man can be a member of his family, an auto club, a discount club, etc....
  3. Every group gathering will enhance the group memory (topo-map memory) by lowering the activation resistance for the group signal. This is called "burn in."
  4. Some topo-maps have much deeper burnt in than others, and they can be activated much easier than others. Smaller the topo-map's size is, easier it can be activated.
  5. Reg2nd-map forms a "small" topo-map network.
  6. Reg2nd-map can link to many other "small" topo-maps networks via a very-alike mechanism and, thus, provides a recall memory mechanism.
  7. Reg3rd-map (or higher order registrations) links all "small" topo-map networks (reg2nd-maps) together.
  8. Topo-map is not a hardware structure in the t-neuron mass. For two identical t-neuron mass (with identical neuronal connections), one might have only 1,000 topo-maps stored while the other has over one trillion topo-maps in it.

For the sake of discussion, I am mapping (projecting) all topo-maps of a t-neuron mass onto a ball surface, depending upon its activation resistance level and its registration relationships. This ball will, then, look like a golf ball. Every point on this golf ball represents one topo-map. The topo-map which has the lowest activation resistance sits on the bottom of the valley of this ball surface. The inside of this ball is an "internal energy wheel" of this t-neuron mass. Obviously, only those topo-maps sit in the bottom of the valley can be reached by this energy wheel directly. Those topo-maps which sits on the hill can only be activated indirectly or by some external energy -- the window signal.

This golf ball is my thinking machine -- with billions (or trillions, ...) topo-maps and with an internal energy wheel. Thinking is, thus, done in the reg2nd and reg3rd sections because reg2nd and reg3rd are converging points of many topo-maps and will have longer "burn in" time and lower activation resistance. Thinking does not rely on any window signal. Although the reg-(beautiful girl) can be recalled, the ws-topo-(beautiful girl) cannot easily be activated by this internal energy.

In a given time, a few points (topo-maps) on this golf ball surface will be lighted up by this internal energy. I call this a "frame" or a "page." In the next time step, it can have a few other points being lighted up, or those old points can begin to grow and to migrate. Thus, thinking can start from some initial "thinking points" which form the first page. The moving of the frames (pages) becomes a "thinking process." With some thinking points and with a thinking process, a few different thinking and reasoning processes can be designed.

Per my design, every "page" will also be registered. That is, an entire thinking process (a book) can be recalled with a much more efficient manner than its first pass creation can be. This "booking mechanism" gives the thinking machine some special properties.
  1. The more a machine thinks, the more efficient it becomes.
  2. With some preferred thinking pathways, a thinking machine could be trapped by those pathways and lose its creativity, for not trying any new pathway.

Now, this t-neuron mass has four sections:
  1. Window signal mapping section -- ws-topo-map
  2. Short term memory section -- reg-map
  3. Long term memory section -- reg2nd-map
  4. Thinking section -- reg3rd-map.
Because the reg3rd-map must also be a part of the long term memory, the section 3 and 4 have some overlaps. That I list them separately is just a design choice.

Before the design of this t-neuron mass intelligence machine, there were two contradictory facts:
  1. Physics laws govern the rise of the entire universe which, of cause, includes intelligence.
  2. The issue of intelligence is beyond the scope of physics.

Now, things have changed. We can soon show that physics laws and the underlying principles of intelligence arise from the same source.

[goto other topics]