A Call-to-Action: Now is the time to Build a Robot with Simulated Consciousness and Natural Language v0.04

 

Introduction

Today, considering the poor success rate of Artificial Intelligence (AI) systems over the past 30 years, very few scientists and computer programmers believe it is possible to build a practical computer based or robot system that is Natural Language (NL) capable, let alone one that simulates human consciousness in any realistic manner. The IBM Big Blue project, the Cyc system that was created by Doug Lenat, the MIT Media Lab robots, and the ASIMO robot from Honda have all advanced AI and robotics, but none have reached even 1% of the capabilities of the Mr. Data character from the Star Trek TV series. Artificial Life (AL) has had some interesting successes in simulating the behaviors of life-forms, however, the author has not found any AL robots that simulate consciousness or goal-directed behavior from the perspective of the simulated life-form to date.

One reason for this state of affairs is that prevailing opinion seems to be that the NL problem is just too complex and to hard to solve in the near future. Due to its mystical trappings, only a few philosophers and scientists have chosen to take the professional risks of even writing about the concept of consciousness, what it is, and how it might work. In another camp there are a growing number of overly optimistic scientists, software engineers, and programmers who think that all barriers to success in the AI field are of our own creation. This latter view is the result of the New Age Kantian thinking of modern Constructivism, which asserts that all we need do to succeed in AI is to think about the problem using a different set of arbitrarily constructed axioms, rather than the traditional ones we have inherited from our ancestors. The Constructivists hold this view because they think reality is subjective and invented by the human mind, not objective and discovered. Here is a quote from one of their websites:

“ Reality is not a discovery, but an invention - a construction based on experience and knowledge. Memory never brings back reality. Memory reconstructs. And each reconstruction changes the original, creating new frames of reference that inevitably grow apart from the truth.

Radical constructivism is an unconventional approach to the problem of knowledge and knowing. Its basic assumption, which is totally independent from the concept of knowledge, is that every individual's knowledge is a unique and inimitable construct that the individual is constantly generating as a closed system in its environment. …

If radical constructivism is right, which we believe, then knowledge and experience are essentially subjective, and although a person may believe that his knowledge and experience are no different from anyone else's, there is no way he can be sure they are the same.” (see reference 8, page 8)

Neither of these views is correct. They are part of a false alternative that is the result of a number of misconceptions about the goal-directed nature of life-forms, consciousness, causality, concept formation, and logical induction.

Consciousness is neither mystical nor inherently subjective . Consciousness is a biological process, not a mystical one. Nor is consciousness an epi-phenomenon, a subjective, non-causal side effect of brain function in animals. Consciousness is a goal-directed, limited, identifiable, relational process , one that only exists in certain life-forms to aid their survival efforts. Consciousness is also causal because it inputs energy from the world and outputs information; it produces memory contents in a continuous series of events. These contents then get used by the brain to enable a life-form to cause changes in the world through its actions. To perform these functions, consciousness, by means of the nervous system in higher animals, transduces the energy of light, sound, pressure, and so on from outside the body of an animal. The nervous system converts the energy into new forms and ultimately into information , then stores it in memory as the content of consciousness inside the life-form. What is outside the life-form is the object of consciousness. After the energy from the object is converted into information (in both analog and symbolic forms), it becomes the subject of that conscious event. If the resulting content corresponds to its object in reality in a one to one relationship, then the content is said to be objective . If it does not, the content is said to be subjective , which means it is either arbitrary (the result of memory manipulation) or a hallucination (an input processing error cause by the limitations of conscious processes). (see references 1,2, and 7)

Consciousness as a process is simply another layer of biological process complexity in the architecture of some kinds of life-forms. It is a process that evolved because it aids in survival. As Ayn Rand put it: “Existence IS Identity. Consciousness IS Identification.” By using the process of consciousness to discover, to identify reality, some animals establish an informational relationship with the world outside themselves that simpler life-forms do not have the capacity to form. By doing so, these animals increase their chances of survival because they can then use the information that this relationship produces to better predict and self-regulate their actions in the world. In this regard, Man is at the top of the pyramid of life. (see reference 7)

Another concept that is often ignored in explanations of consciousness is causality . This occurs because most scientists implicitly believe two things: First, that consciousness is either mystical or non-causal (they take one side or the other of the false alternative indicated above), and Second, they think of causality as “billiard ball,” action-reaction kind of events, not as an identity-action relationship . Granted, causality is a complex topic, but if one thinks about it, what objects do depends on what they are . For example, drop and egg and a helium balloon a meter off the floor, and while it is true an event occurs, there is really more going on. Each of the objects involved in the event is an identity. In other words, each object (including the earth) is an integrated set of features or characteristics (size, shape, weight, mass, air envelope, buoyancy, position, color, and so on). When the balloon and the egg are dropped, what is really going on is that the identities of the objects interact . The egg falls and smashes on the floor while the balloon floats. Why, because of their identities in relation to those of earth, including the earth's characteristics of gravity and air. The balloon is buoyant in air, and the egg is not, so the identity of the egg reacts differently with the earth than the balloon does. Causality is the interaction of the objects' identities , not merely an event.

Observational and inductive examples like this imply an important principle: The action capacities of objects depend on their identity, or another way of putting it is that the identity of an object determines its action capacity . This principle has an important implication for AI and AL: Systems that can change their own identities also change their own action capacities at the same time . Life-forms do this all the time. It is called learning. AI and AL systems can simulate learning, but only if their designers understand the nature of consciousness and causality, and how these processes operate in goal-directed biological life.

The other important misconception many scientists have is this: Every physical scientist knows that to attempt astro-physics without calculus leads to epi-cycles, not orbits. As Newton and others discovered, to plot the paths of the planets, one needs calculus. But calculus depends on algebra, and the attempt to use calculus without it is tantamount to trying to use modern software applications without an operating system. Knowledge is layered, and one needs all the layers to make it work. Attempting to build functional AI systems without a scientific, methodical, mathematical theory of concept formation is an analogous case because much of our knowledge, like many technology architectures, is layered.

Much of our knowledge consists of NL sentences like the ones you are reading. The sentences are common language forms of premises that depend on and communicate logical propositions about some subject. The propositions consist of chains of concepts. The bottom layer and source of all of this content is our sense perceptions of the world, of reality. Everyone uses all these layers to function in their daily lives, and we all use concepts almost without thinking about it constantly, but where do concepts come from? It is not enough to just say: “Well, they are here, so who cares?” At least it is not if one's objective is to build an NL AI system or robot. Most scientists probably do not believe that concepts are “mystical or merely emotive and intuited.” AI has almost exclusively tried the “arbitrary, subjective, pragmatic, make them up as we go definition approach” for the past 30 years without success. So what alternative remains? The answer is to form objective concepts scientifically and methodically using psychological measurements implicit in the human perceptual system. People can use these implicit measurements to extract the identity of the objects we observe. We can develop robots that can do so too. In attempting to simulate intelligence as AI systems and some robots do, the perceptual identification and concept formation layers cannot be left out. The reason is the same as for the calculus and operating system examples above. In this case, NL depends on valid, objective concepts and cannot function without them. So does logical induction to generate premises from observations of reality, but that is another story. (see reference 2)

The whole point of AI and AL is to simulate the capabilities of life-forms using various kinds of technology. While it is true we are “not quite there yet” in terms of solving the NL problem, with a design architecture based on the foregoing explanation, the author believes we are close, probably only 2-5 years of concentrated effort away by a strongly motivated workgroup (sooner for a convincing simulated “block world” type demonstration). The following are the key requirements the author believes must be met to make a convincing demonstration of a practical consciousness simulator that can use simple, one phrase NL sentences to identify its world and itself, as well as to communicate with people effectively:

1) A simulation of consciousness must operate as a self-sustaining, self-regulating, and goal-directed system in a manner that includes the management of its own identity, so it can regulate its own action capacity utilizing its own observations of reality.

2) The simulation must perceive the human-scale world of ordinary objects and their actions as ordinary people do in order to make observations of reality that approximately match human observations.

3) The simulation must be able to use simulated volition (free will or the capability of making choices) as part of its procedure for forming objective  concepts. It must do so using its own observations and Ayn Rand's quasi-mathematical concept formation method, and it must complete the concept formation process by symbolizing this new data type with NL words that are provided by a human tutor.

4) Using more observations and the same method described in requirement 3 above, the simulation must be able to volitionally form abstractions from abstractions (more abstract concepts from simpler ones) resulting in unbroken chains of hierarchically and contextually organized concepts , every one of which is derived from perceptual observations (which means none are arbitrary = subjective). Then the system must validate these new abstractions using other means such as deduction, reduction, and integration with other fields of knowledge to eliminate contractions from the system with the help of a human tutor.

5) The simulation must volitionally use the resulting NL/conceptual system and additional observations of reality to form generalizations of causal and other relationships to generate premises (premises = sentences = complete thoughts) that identify the causal sequences the simulation observes in reality. Then the system must validate them using other non-inductive means such as deduction, reduction, mathematics, and integration with other fields of knowledge to eliminate contractions from the system.

6) The simulation must volitionally use this entire objective, ever-growing, NL based system of concepts and generalizations, along with the constant input of more observations of reality, in order to self-regulate its own identity. It must do so in order to increase its own range of choices, which will, if correctly made, increase its control over reality. And it must do all of this to promote its own survival to cause its own future simulated happiness.

The requirements just outlined are straightforward and well within the reach of current technology, at least in their simplest form, and can all be made part of a robot control system. The theoretical basis for all of these requirements, their interdependencies, and a detailed explanation of how to build a consciousness and NL simulation system (including demo code) are all explained in the references for this document in detail, so the reader is encouraged to look to these references to find answers to the many questions that are probably starting to come to mind.

The remainder of this document will look at each of the requirements listed above more closely and add additional explanation. The document will also briefly consider the market needs for each of these functions, the available technologies to implement them, the estimated time to develop reasonable demonstration applications, and the approximate cost to do so.

Now let us consider each of these six requirements in more detail.

1) Life-Like Simulation: A simulation of consciousness must operate as a self-sustaining, self-regulating, and goal-directed system in a manner that includes the management of its own identity, so it can regulate its own action capacity utilizing its own observations of reality

    A) A bit more explanation: Simulating goal-directed behavior using computer systems is not difficult. There are many computer based simulators in existence already that simulate individual plants and animals as experimental subjects for biological study, pet robots, and life-like computer animations (see references 5 and 6). There are also simulators that simulate groups of organisms for ecological research purposes and some that even simulate entire ecological systems (see reference 5). These systems are constantly becoming more sophisticated. What none of these state of the art systems does, however, is simulate goal-directed behavior from the perspective of the life-form being simulated. Virtually all such systems the author has found to date are designed for human purposes, not the purposes of the simulated life-form itself. One of the reasons for this is the mistaken view that life-forms are mechanistic, that life-forms are mere machines. In fact, their operation is more causally complex than machines, and grasping this important difference is crucial to successfully simulating life-forms. (see reference 1, chapter 2 and reference 7 for details) The only change that is needed to satisfy the requirement of taking the more complex causality of life into account is for a simulator program to be written that enables the simulated life-form to operate for its own ends , its own primary goals and perspective, rather than those of some human designer, experimenter, or customer. To do so is really quite simple: A programmer needs only to imagine life from the perspective of the simulated life-form and write code that enables the simulated life form to live for itself first, to make its own survival the primary objective, and to put human agendas second : Call such a system an Ego-centric Robot Control TM program*(see reference page) . If one observes life-forms in reality, one will soon discover that this is what they actually do with few exceptions. Of course, one disadvantage is that such robots would have to be trained to help accomplish human goals as horses and dogs are trained to do, but with the advantage that when operating on their own, they could make independent decisions and take independent action. Ego-centric robots would, of course, make “mistakes” from time to time, but then nothing is for free. The new and most important idea here is that while life processes are based on mechanistic chemistry, they lead to emergent properties that are more causally complex than non-living processes. Life processes are self-generated, self-maintained, and self-regulating in a manner that non-living processes are not, as the following diagram shows. Dead life-forms cannot act. (Dead life-forms cannot act. Survival is a complex causal process: Life causes both its own continuation, and as long as life actions continue, it can cause other actions that may not necessarily be survival related. The following image depicts this complex form of causality graphically. The goal of survival must be met on each cycle for the process to continue, and as you can see, it is very different from the way mechanisms work.

    (See reference 3 in Appendix A Section 1.2.1 of our book or reference 7 below.)

    B) Need for this Functionality: The robotics industry is crying for systems with the capacity for independent action for any applications that cannot be done well by remote control such as for space and undersea missions or bomb disarming and disposal at long distance. It is well known that the greatest short-coming of state of the art robots is their limited ability to act independently in situations such as these examples imply. While there are many other applications, applications for robotics such as these are probably the biggest and will be the fastest to grow once this technology becomes widely available.

    C) Available Technology: State of the art simulation technology with the addition of Ego-centric Robot Control could easily be produced with modified versions of today's robot software development environments. It is more a matter of how the code is designed, written, and used, rather than changes to the technology itself. It should be noted however, that ego-centric robots will need to perceive the world as explained in requirement two below in order to work properly. The modifications necessary to existing robot control code development environments will be to include options for writing goal-directed behavior code, options they currently do not have. This is necessary so the robots can be goal-directed from their own perspective, so the robots can set and accomplish their own goals , rather than those of a programmer or user. The description, explanation, and examples of how this can be done are contained in chapters 3, 4, and 5 of reference 1.

    D) Demonstration System Project Plan: The plan to develop an Ego-centric Robot Control system will need to include at least the following milestones:

      • i. A design that is based on and effectively captures and simulates the complexity of teleological (goal-directed) causation rather than the simpler mechanistic, “billiard ball-type,” action-reaction events that most state of the art systems use. The simulated life-forms in this system must be self-generating, self-sustaining, self-regulating, and self-modifying. The design must also contain an accurate and Ego-centric simulated pleasure/pain system that holds the simulated life-form's “life” as its highest value and standard, and then measures perceived events against that standard when evaluating its observations of reality.

      • ii. Once an appropriate design has been engineered, then the usual Alpha, Beta, and Final developmental milestones must be reached to write and test the simulation code. Care must be taken during these steps that the code will be a good foundation for and smoothly integrate with the code for the other simulation requirements.

    E) Estimated Man-Years to Develop a Practical, “block world” Demo: 1-2

    F) Estimated Cost: : $1-200,000.00

 

2) Human Scale Sense Perception: The simulation must perceive the human-scale world of ordinary objects and their actions as ordinary people do in order to make observations of reality that approximately match human observations.

A) A bit more explanation: Simulation of human scale sense perception is not what state of the art robots and computer systems do. Modern robots and computer systems are machines which “see” the world as bitmaps. All their subsequent mechanistic processing uses bitmaps as data. People are goal-directed and see the world as objects that interact in various ways as part of an integrated scene. People are much more highly capable in a wide range of conditions than any technological creation humans have built so far. This is because the visual and other parts of the neo-cortex in the animal and human brains works in very different ways than state of the art computers and robots do. Animals and people are not machines, and they do not see with bitmaps. (See reference 4 for details) The key things to grasp about biological sense perception is that it is the source of all the content of animal and human consciousness, and that sense perception does not recreate a 3D image of reality inside the brain. The sensory-perceptual system transduces energy that has reached its sensors from reality. Light energy that reaches the eyes or sound energy that reaches the ears is transduced into neural-electrical energy and transmitted to the neo-cortex in the brain. There it is further processed and some of it is transduced into chemical energy when it is stored as memories in the synapses of neurons. These are some of the goal-directed, biological processes of consciousness, the objective of which is to identify reality to aid in the survival of the animal or person. As Dr. Harry Binswanger has pointed out, “the brain is a difference detector.” The goal-directed process of detecting the differences in reality produces patterns in neuron firing and molecular changes in synapses. It is these patterns that are the content our minds are aware of when we perceive the world. The beginning of seeing (and all other perception) is sensing and detecting differences in reality. That is the base level content and the foundation of all knowledge . Everything else, such as the cognitive function of pattern recognition, is abstracted from the differences that have been detected by the conscious or subconscious processes that are layered on top. But it is the differences , and the various forms in which these differences can exist, that consciousness uses to identify the world outside---and to identify itself. Without these differences, there would be no patterns to identify and recognize .And while such simulations can never equal biology (precisely because they are not alive), they can mimic some biological processes, behaviors, and causation, provided they are designed with the appropriate identity. The layered model below shows this system design in the form of a table. Note: The functions in each layer of the table are analogous, not equivalent.

Biological Life-Forms

Digital Life-Forms

Layer 7

Conceptual Consciousness (Reason)

Simulated Conceptual Consciousness

Layer 6

Perceptual Consciousness

Simulated Perceptual Consciousness

Layer 5

Goal-directed Cellular Processes

Simulated Goal-Directed Behavior

Layer 4

Mechanistic Cellular Processes

Digital Life-Form Simulation Program

Layer 3

RNA, Protein, ATP Synthesis

Object-Oriented Prog. Environment

Layer 2

DNA Processes

Computer Operating System

Layer 1

Electro-chemical, Physical Processes

Computer Hardware

    B) Need for this Functionality: Any application in computer vision systems and robotics that requires processing pictures of what animals or humans see will benefit from sense perception technology that works more like biological visual processing by being modeled after the way the cortex of real brains work.

    C) Available Technology: None at present, but software tools to help develop products that process visual and other sensory information for pattern recognition is in development by Numenta, Inc.

    D) Demonstration System Project Plan: The plan to develop a human-like perceptual system involves working in a customer/vendor relationship with Numenta, Inc. to see if their products can be used as is or in modified form to produce the information required by a simulated consciousness to identify reality. For this process to succeed, Numenta products must:

      • i. Use off-the-shelf sensor technology

      • ii. Process the output of sensors such that differences in reality at the human scale that are normally observed by people looking at the world can be easily detected and transferred inside a simulation system

      • iii. That once inside the system, identity information is conserved as it is transferred to the simulated neo-cortex processors

      • iv. That the output of the information in the system is “chunked” as objects that correspond to the objects that people normally see

      • v. That the informational objects thus detected can be automatically identified by a list of features or characteristics, such as color, shape, location, weight, hardness, texture, and so on

      • vi. That the features or characteristics are output in the form of quality (color, length, speed) and quantity pairs, where the quantity is a measure of the quality

    E) Estimated Man-Years to Develop a Practical Demo: 2-3

    F) Estimated Cost: $2-300,000.00

 

3) Concept Formation: The simulation must be able to use simulated volition (free will or the capability of making choices) as part of its procedure for forming objective  concepts. It must do so using its own observations and Ayn Rand's quasi-mathematical concept formation method, and it must complete the concept formation process by symbolizing this new data type with NL words that are provided by a human tutor. (see reference 1, chapter 5 for details on what the method is and how it works)

    A) A bit more explanation: The simulation of the human concept formation process is a complex and sometimes controversial subject. In the entire history of epistemology (the study of human knowledge), all the ways of forming concepts can be reduced to the following three:

    • i. Intrinsicism: The use of intuition, feelings, or some other subconscious means to produce concepts.

    • ii. Subjectivism: The use of purely arbitrary selection of the content and definitions of concepts (usually for pragmatic reasons).

    • iii. An Objective Method : This approach was first identified by Ayn Rand in the early 1960s and published in her book Introduction to Objectivist Epistemology in 1967 (see reference 9). Her procedure is the only method-based approach the author has ever found for human concept formation that covers the entire universe of human knowledge. Given that concepts are the building blocks of all knowledge humans retain using symbols, the method one uses to form concepts is more fundamental than the scientific method itself. The scientific method, after all, depends on and uses concepts for its own definition. Rand's concept formation method is a quasi-mathematical procedure: It requires all concepts be based on sense perception and grounded in a specific context and range of comparative, psychological measurements (this looks bigger than that, but smaller than something else). Concepts can also be formed from other simpler concepts leading to unbroken, hierarchical chains of concepts that connect to sense perceptions of reality at their base (observations), and these more abstract concepts must then be validated by integration both hierarchically and with other, known valid concepts in two or more diverse fields of knowledge.

    Rand's method requires people use volition (choice) to form concepts because there are a potentially endless number of concepts that could be formed, and hence choices must be made as to which ones are necessary and when. In order for robots to simulate an objective concept formation method, they will also have to simulate making choices. While the details of the method and how to simulate volition are outside the scope of this paper, it is the author's view that Rand's method is the only one of the three ways of producing concepts that lends itself to building a computer/robot simulated life-form attempting to classify its world for its own sake , not the agenda of a human programmer. Human intuition and arbitrary concept definition by human programmers have both been used nearly exclusively in all state of the art AI systems to date. No self-respecting software engineer would consider using abstract, high level software application variables or instructions that are not logically connected to a development environment being used, the appropriate operating system, and the appropriate computer hardware. Why should human concepts be any different.

    B) Need for this Functionality: The need for concept formation by simulated life-forms is the same as the need of it for humans: The naming of objects and their actions to simplify and speed identification . Doing so reduces both storage and processing needs. A concept formed by Rand 's quasi-mathematical method is a new data-type that is open-ended and timeless. The concept “man,” for example, enables that single symbol to be used to represent all the people who ever have lived, live now, and ever will live, and to do so with one word of only three letters. Ayn Rand calls this “unit economy.” The application of this technology in robotics and other areas could save enormous amounts of storage and processing resources. (see reference 2 for a short explanation of the concept formation process and reference 1 chapters 4 and 5 for a detailed explanation of how to simulate volition and concept formation)

    C) Available Technology: There is currently no product known to the author that contains and uses this technology, however, in the early 1980s the author wrote and tested a simple, proof of principle program that successfully classified closed geometric shapes such as triangles, circles, and rectangles using the method. (see reference 1, chapter 5 for an update discussion on this process)

    D) Demonstration System Project Plan: The plan to develop a simulated life-form with the capability to form concepts on its own depends on the development and testing of requirements 1 and 2 above, however, a much more robust demonstration system in a simulated “block world” could be developed quite easily using the object-oriented code in reference 1, chapter 3 as a model and starting point. A more sophisticated design could then be developed for the new demonstration system as a convincing proof of principle program. When understood from a methodical point of view, concept formation is a straight forward process, especially at the base level of forming concepts of perceived objects (things in the world such as tables, rocks, cars, houses, trees, people, and so on. This assumes, of course, that one has a simulated sense perception system that can be trained to extract the identities of these objects from the output sensors such as digital TV cameras and microphones, as described in requirement 2 above.

    E) Estimated Man-Years to Develop a Practical Demo: 1-2

    F) Estimated Cost: $1-200,000.00

 

4) Abstraction from Abstractions: Using more observations and the same method describe in requirement 3 above, the simulation must be able to volitionally form abstractions from abstractions (more abstract concepts from simpler ones) resulting in unbroken chains of hierarchically and contextually organized concepts , every one of which is derived from perceptual observations (which means none are arbitrary = subjective). Then the system must validate these new abstractions using other means such as deduction, reduction, and integration with other fields of knowledge to eliminate contractions from the system with the help of a human tutor.

A) A bit more explanation: Base or first-level concepts, that is, concepts of perceived objects are very basic and simple. Many of them don't even have verbal definitions, but are retained visually using “ostensive” definitions (by simply pointing to examples). Humans who have learned to use language form this type of concept all the time and in many cases without even realizing it. Second level and higher abstractions, however, are more complex because they depend on earlier formed, simpler concepts and form complex hierarchies in the human system of knowledge and require verbal definitions just like computer software variables and instructions do. For example, the concepts “table,” “chair,” “couch,” “bed,” “dresser” are first level concepts. The objects these concepts subsume are all directly observe-able. However, the second level abstraction “furniture” is not. There is no single object: “furniture,” only tables, chairs, and so on. The same is true for higher level abstractions: There is no archetypal object that one can observe to form the abstract concept “object.” Actions and relationships are also abstractions from abstractions. In fact, while actions can be perceived as motion, relationships are completely invisible. In the sentence “The dog ran across the rocks.” the only perceive-able objects are the moving dog and the rocks. The action of “running” and the spacial relationships of the objects' positions are abstracted from the objects and the perceptual scene (the context) in which they occur. This relational information is implicit in the scene and extracted by processing in the neo-cortex of our brains as part of our perceptual system. The neo-cortex automatically extracts the raw information we need and then we can choose to use it to conceptualize what is implicit in the scene. It is up to us to make sure we use this information correctly. We must make the correct choices in this process so our concepts are connected together in unbroken chains by valid concept definitions. It is the definitions that connect the concepts to each other and ultimately to our observations of the scene itself, the context from which the concepts originated. This must be done in order to conserve their meaning . In human knowledge, it is unbroken chains of concepts that are tied to reality, to specific sense perceptions that contain the meaning of NL sentences. It is these unbroken chains that make NL technology work. A human thinker depends on these unbroken conceptual chains, just as a software engineer depends on the built-in logical connections in a computer software development environment to keep his/her code connected to the hardware that processes it. In other words, if we are to simulate NL in robots, we need to connect the NL words to reality to make sure they mean something to the robot .

B) Need for this Functionality: While robots that could learn to classify and name objects in their world as members of classes would certainly be an advance in state of the art technology, it would not be nearly as valuable as robots that could, with the help of human tutors, identify higher level abstractions like concepts such as “object,” “action,” “above,” “between,” and so on. Doing so is one of the key steps to being able to perform independent action or take complex verbal instructions from people, something not possible to state of the art robots.

C) Available Technology: There are currently no products available that contain the technology to form abstract concepts, however, once a the system is developed and tested to form first level concepts, forming abstractions of at least the next few levels will be more straight forward. The procedure to do so simply leverages and further processes the data produced by simulated sense perception and in first level concepts. The same concept formation processes are applied (along with new data) to build conceptual hierarchies of unbroken conceptual chains, and it is these chains that conserve the meaning of the words that symbolize them by keeping the words connected to the actual objects in reality used to form the concepts in the first place. The technology to form higher level abstractions will be a reasonable next step from the development of the demonstration program described in 3 above.

D) Demonstration System Project Plan: The plan to develop a simulated life-form with the capability to form abstract concepts depends on the development and testing of requirement 3 above. Once that layer of the system is operating and a few hundred first level concepts have been formed, a robust demonstration system can be developed using the code that will be produced by accomplishing requirement 3 as a model. A more sophisticated design can then be developed from this demonstration system for both levels of concept formation as described in detail in reference 1, chapter 5. It is essentially another aspect of the same process, and the re-processing of some of the same data used earlier.

E) Estimated Man-Years to Develop a Practical Demo: 1

F) Estimated Cost: $100,000.00

 

5) Logical Induction to NL Generate Sentences: The simulation must volitionally use the resulting NL/conceptual system and additional observations of reality to form generalizations of causal and other relationships to generate premises from its own perspective (premises = sentences = complete thoughts) that identify the causal sequences the simulation observes in reality. Then the system must validate them using other non-inductive means such as deduction, reduction, mathematics, and integration with other fields of knowledge to eliminate contractions from the system.

A) A bit more explanation: Just as the author does not believe concepts are intrinsic to either reality or our brains but a relational combination between the two, the same can be said for the genesis of premises (the sentences of NL). Logical induction is not about statistics, it is about premise generation . Concept formation generates the system of symbols people use to classify and identify reality. Once a sufficient number of concepts have been formed, people routinely use concepts for the identification of objects they perceive in reality. Consider the example used above: “The dog ran across the rocks.” Given that scene and the necessary concepts, the human neo-cortex encodes that sentence by simply using the union of the concepts that get triggered by the identification of the scene. It can do so because the percepts of the objects observed in the scene fit the measurements of the pre-existing concepts of “dog,” “rocks,” and “running.” Those concepts produce a union in our neo-cortex as they are identified subconsciously. Since each concept is associated with a word, those three words are also activated in the subconscious, and come to the conscious mind as the answer to the implicit question: “What is that?” Our previous learning of some NL applies the proper syntax, and we have a premise, which is also an NL sentence or complete thought. Depending on which language we speak the words and their order may vary, however the meaning will be the same (or nearly so). Obviously, actual practice is somewhat more complex, but this simple example shows how the process works. It really is not “rocket science,” once the foundation has been put in place with the lower layers of the system as per requirements 1-4 above. (see reference 2 for a summary and reference 1, chapter 5 for complete details)

B) Need for this Functionality: NL processing is a natural consequence of a fully developed conceptual system, just as computer programs are a natural consequence of all the pieces provided by programming development environments. In this case, just as NL words must be supplied by a tutor, the syntax and grammar of their use must be as well, which can be as different as they are for computer languages. What is missing from the state of the art is the counter-part in our understanding of the human mind that represents the “programming environment” in software engineering. That counter-part is what the methodical approach to concept formation and inductive generalization produces. NL follows naturally (no pun intended!). The bottom line is that few people would argue there is no need for NL technology in robotics. The problem has been to make it work. The technology of objectively formed concepts and inductive generalization from observation of reality provides the solution.

C) Available Technology: There are currently no products available that contain this technology. However, as is the case with abstraction from abstractions, once a conceptual system is in place, the next level of processing builds on the earlier ones. In this case, NL processing for both encoding and decoding sentences depends on all the lower layers of processing described in the previous requirement sections, but mainly on simulated perception, first-level concept formation, and abstraction from abstractions. The latter is especially important because that layer produces the validated and unbroken conceptual chains that can be traced up or down to formulate or determine the meaning of words and sentences.

D) Demonstration System Project Plan: The plan depends largely on the development and testing of the lower processing layers described above. While this layer of processing is very complex and difficult, the author believes that it can be done with reasonable ease for simple, subject-verb-object sentences of 4-6 words. (At this time it is impossible to estimate how to build a system to encode and decode complex sentences such as those you are reading in this document.) The author's confidence about simple sentences, however, comes from the ability to do thought experiments with a number of examples using closed geometric figures that easily produce straight forward results. (see reference 1, chapter 5 for complete details)

E) Estimated Man-Years to Develop a Practical Demo: 3-4

F) Estimated Cost: $3-400,000.00

 

6) Simulated Self Consciousness: The simulation must volitionally use this entire objective, ever-growing, NL based system of concepts and generalizations, along with the constant input of more observations of reality, in order to self-regulate its own identity. It must do so in order to increase its own range of choices, which will, if correctly made, will increase its control over reality. And it must do all of this to promote its own survival to cause its own future simulated happiness.

A) A bit more explanation: There have been numerous previous attempts using AI systems to simulate consciousness, and most of them have not worked very well. These attempts have shown that this problem must be tackled as a whole, not piecemeal, that the nature of consciousness cannot be ignored, nor can the goal-directed nature of biology. Life-forms are complex, integrated, goal-directed, self-regulating systems that hold the continuation of life processes, their own survival, as their highest value. (see reference 1, chapter 2 and reference 7 for details) This fact is the driving force, the motivation to take action on the part of life-forms. Simulated life-forms designed to mimic their real counter-parts will act in a similar fashion. This conclusion leads to some interesting consequences:

      • i. As with any potentially dangerous equipment, robots based on the technology described in this document will need many built-in safe guards to make sure they do not harm people.

      • ii. Robots that simulate animal and human consciousness will need to be trained, rather than programmed. Sooner or later, some of them will make mistakes with negative consequences.

      • iii. Robots that simulate human consciousness will have unpredictable capabilities just like people do. As pointed out in the introduction, action capacity is a consequence of identity: A hammer drives nails because of its design. An egg smashes if dropped because of its mass and the make up of its shell, whereas a helium balloon floats in the air. Some people are overweight and others are athletic because of what they eat and how much they exercise. Some people become laborers and others professionals because of what they put in their brains. In other words, choices lead to changes in identity, which in turn lead to changes in action capacity. The same will be true for robots designed using the technology described in this paper. The difference is that robots can operate 24 by 7 indefinitely, and will have access to the Internet. Consider the possibilities. They must be designed carefully.

    B) Need for this functionality: There will be some people who will argue that we do not need this technology, just like some people argue we do not need nuclear technology, nanotechnology, genetic engineering, space travel, and so on. The arguments are the same in each case, and the benefits to human survival out-weigh the risks. The benefits of having robots with even 10% of the capability of Mr Data from the Star Trek TV series are obvious.

    C) Available Technology: There are currently no products available that contain this technology. And frankly, the author does not claim to know what the limits of the capabilities the technology described in this document are or how long it will take to reach them. That will depend on the operational success of the system layers described in sections 1-5 above, funding, and the motivation of the design teams. It can be said that if the technology in those processing layers functions reasonably well, that the goal of reaching 10% of the capability of Mr. Data is attainable. Beyond that is unknown. Building the system described in this document is not an issue of hardware processing speed or available memory. It is simply an issue of designing a complex software system that mimics goal-directed behavior from an ego-centric perspective of the system itself, simulating sense perception with a neo-cortex along the lines that the author has proposed, simulating the formation of concepts using the objective procedures and quasi-mathematical approach identified by Ayn Rand, forming and validating abstractions from abstractions using a slightly more complex form of the same concept formation technology, and using logical induction to encode sentences within the conceptual system as the basis for NL processing. This project is much less ambitious than building a Saturn 5 rocket and going to the moon. Its prototype can probably be built on any reasonably powerful personal computer.

    D) Demonstration System Project Plan: The plan to develop and train a robotic system with the capacity for simulating self-consciousness depends largely on the development and testing of the lower processing layers described in requirements 1-5 above. While the layer of processing described in layer 6 is the most complex and difficult, the author believes that it can be done, provide the lower level processing can be made to work effectively. No definitive answer can be provided as to development time and cost until further development work has been done and evaluated. Only approximate estimates can be made at this point.

    E) Estimated Man-Years to Develop a Practical Demo: 5-10

    F) Estimated Cost: $5-900,000.00


Conclusion

Based on the foregoing, it seems clear to the author that the technology is at hand to build some level of robot control system that simulates consciousness and the goal-directed behavior exhibited by life-forms. It should be at least possible to build a robot that can classify and name the objects it perceives using concepts in 2 years and probably one that uses simple NL sentences within about 3-5 years. The estimated development cost would be approximately $1.2 million dollars for the latter. But a convincing demonstration using a block world that would show the value of sense perception to identify reality and concept formation to help control robots would take less time and likely could be developed for less than $1 million dollars.

The break-through by Jeff Hawkins in identifying the operation of the neo-cortex and the way neurons apparently process sense perception is an important step on its own. It is the basis for the content for the rest of the simulation system.

The software tools Jeff's company is developing will make building a robot perceptual system much easier, assuming the technology being developed can be used to develop code that can output the identities of the objects its sensors are aimed at. Only careful design, experimentation, and testing will verify that. If so, when put together with the technology of the quasi-mathematical and objective concept formation procedure and other technologies described in this document, an entirely new path to success in robotics opens before us.

It is time to follow that path.

 

References

  1. How to Simulate Consciousness Using a Computer System - Gregory J. Czora, Copyright 2001, published only on the Internet, http://www.blueoakmountaintech.com/DLF_Book.html/Cover.html
  2. An Inductive, Biological Approach to NL and Math – Gregory J. Czora, 2005*
  3. Invariant visual representation by single neurons in the human brain , Quiroga, Reddy, Kreiman, Koch, and Fried, Nature:Vol 435|23 June 2005 |doi:10.1038/nature03687
  4. On Intelligence - Jeff Hawkins, Owl Books, 2004, ISBN# 0-8050-7853-3
  5. Heleno, et al, Artifical Animals in Virtual Ecosystems , published in Computer Networks and ISDN Systems, Volume 30, Issues 20-21, November 1998, pages 1923-1932
  6. Yokoo, et al, United States Patent 6,449,518, Storage medium, robot, information processing device and electronic pet system
  7. The Biological Basis of Teleological Concepts - Dr. Harry Binswanger, Ayn Rand Institute Press, 1990, ISBN# 0-9625336-0-2
  8. PIBOT VS TURING (ACTIII) – Rick Genial, September 2005, page 8 http://www.pibot.com/tiki/tikiread_article.php?articleId=3&PHPSESSID=636acb92718421082183167ea104ead2
  9. Introduction to Objectivist Epistemology - Ayn Rand, Meridian , Expanded Second Edition, 1990, ISBN# 0-453-00724-4
  • *Ego-Centric Robot Control is a trademark of Blue Oak Mountain Technologies, Inc.

 

Welcome Page | About Us | Testimonials | Services | FAQs