Why a philosophical theory of the consciousness is necessary
Exemplified by the case of intelligent agents which can learn and speak any natural language
AUTHOR: Dr. Gerd Döben-Henisch
FIRST DATE: Nov 21, 1995
DATE of LAST CHANGE: Nov 21, 1995
TRANSLATED by: Margret HAASE and Ed HARRIS
UNFULFILLED WISHES
Let the vision of Nicholas NEGROPONTE serve as the starting-point of our considerations. In his book "Being digital" (1995) he states that his dream of the ideal interface would be "computers" that are human-like (p.101). Such a human-like interface would have, among other qualities, the ability to communicate and converse with people in a human manner. This also implies, for example, that such an interface could really understand and as well produce language that is suitable to a context. It further implies that such a human-like interface could accept and independantly carry-out commands. Maybe it could even activate certain functions on its own spontaniously.
NEGROPONTE is certainly realist enough to know that even today there is no computer that understands language as we do. Until now, only in the restricted field of language recognition certain success has been acieved. What language recognition means is when acoustic sounds are projected onto character strings of a certain language. All the proccesses used today work with purely statistical methods. Although the methods use phonetic and morphologic rules to a certain extent, more or less, none of them refer, in a mentionable sense, to something like "meaning" or "understanding".
At this time supposedly the most capable and efficient language recognition system, which in September, 1995, I was able to test, was the "Listening Typewriter", produced by Siemens-Nixdorf, developed under the leadership of Dr. Proff. It`s based based on the hardware of a standard PC with a Pentium processor enhanced through a cartridge with a special neuro processor. The "Listening Typewriter" recognizes, independant of who is speaking, fluent speech in real time.
UNSOLVED PROBLEMS
Since the 1980`s under the catagory of "intelligent agents" ( i..e., software agents and robotic agents ) there is a rapidly expanding branch in artificial intelligence research which is aimed at developing machines and programs, with today`s available technology, which perform more and more intelligent human skills under various conditions. The interface mentioned by NEGROPONTE, appears here -if at all- only as a fantastic special case in the distant future. It is obvious that in this field of research the ability to understand language, not to mention language acquistion, by intelligent agents, until now, hasn`t even reached the point of discussion. In the state-of-the-art survey "Agent Theories, Architectures, and Language," by Michael J. WOOLDRIDGE and Nicholas R. JENNINGS (1994), we find the attempt to characterize agents as "intentional systems" which have knowledge, beliefs and wishes at their disposal, but the specific topic of language understanding is not even mentioned.
A cause for this silence is surely the lack of a useful theory of language understanding and language acquisition.
In his knowledgable review article, "Knowledge Representation and Natural Language Understanding" (1993), Gerard SABAH states that (1) the syntax needs the semantics (p. 165 & following ), (2) there is no one-to-one projection between syntax and semantics (p.159) and (3) it is the task of pragmatics to describe which principles are at work when a listener catagorizes the actual sense and the actual reference of an expression (p.169). But, he does not yet see how the actual theories of pragmatics ( he mentions GRICE, SPERBER /WILSON, ROULET) could be translated into executable programs (p.171).
In a recent review of automatic translation systems, Uta SEEWALD, (SEEWALD 1995) found out that all translation systems still need post-processing by a human translator. This effort comprises between 20% and 60% of the total translation work. The encoding of the meaning, which plays a central role in the frame of translation work, "is still an open unsolved problem," as she carefully puts it (p.96), and : " A full automatic machine translation, that doesn` t need post-processing by a human being, is a projection into the future" (p. 102).
These difficulties are not surprising.
EVENTS IN THE MIND OF THE SPEAKER
We would like to examine a very simple example in an unscientific manner.
Let`s take for granted that while I`m writing this article my telephone rings and a voice at the other end of the line states:" Mr. Döben-Henisch?, Please come to the main entrance of the institute. Mr. Meyer is waiting for you there."
First of all, this experience functions as an audio experience, which everyone that understands the German language, would unambiguously " classify as a statement of language". The speaker of this statement- a woman`s voice, recognizable as the voice of one of our secretaries- who, at the time of the statement, is not bodily present in my room. Some of the mentioned objects like, "main entrance" and "Mr. Meyer" are also not present in the situation. Nevertheless, I`m in the position to assign to this audio experience- casually formulated- "a purely imaginary situation" in my mind for which there is no direct equivalent in the actual situation of expression. On the basis of spoken audio experience we can imagine subjects, relationships between these subjects, actions, diverse activities, without which the contents of the imagination, in the immediate statement situation, must be somewhat objective.
This exemplary description of the everyday spoken meaning is- as an unscientific representation- imperatively a little vague. However, it does bring two principal facts to our attention which point the way to further considerations:
- Linguistic meaning and the understanding tied to it has something to do with processes that occur in the mind of the speaker/listener. These processes must not necessarily lead to observable behavior.
- Every speaker/listener has experiential access to his language understanding processes. That doesn` t necessarily mean that all factors relevant for the totality of events must be comprhended.
If one accepts this working hypothesis, one finds strong arguments why many, not to say the overwhelming majority, of today` s assumptions on language understaning processes do not work.
THE EXAMPLE OF EXPERIMENTAL PSYCHOLOGY
Experimental psychology delivers an impressive example of the inadequacy of the treatment of meaning-relevant phenomena. Here, the methodical difficulties are typical for a variety of other disciplines which today deal with human behavior in the widest sense.
Wilhelm WUNDT (1832-1920), the founder of experimental psychology, accepted experiences as a genuine data source, in the form of introspection, for psychological research. Moreover, he was aware that only a small part of the occurences, which are ascertainable through introspection, let themselves be "objectified" through behavior based experiments. almost none of the higher, more complex, introspective ascertainable processes let themselves be catagorized as direct behavior equivalents.
As experimental psychology based on WUNDT attempted to allign itself with such successfull empirical paradigms, as can be found in physics, step by step, it phased out introspection as a data inquiry method. John Broader WATSON` s behavioralist manifest of 1913 " Psychology as the Behaviorist Views It," pointed out what had already been prepared by numerous authors in one form or another. " Introspection" as a method was banded; every reference to "consciousness" was rejected. The modern experimental psychologist sees himself now also as an observer who regards systems as input-output systems and examines them exclusively " from the outside" . Statements about the subjects of reseach are only allowed insofar as they are based in observable events. The big players of the New- Behaviorism were Edward C. TOLMAN, Clark C. HULL and B.F. SKINNER.
If one sticks to the previously formulated working hypothesis, that the things relevant to understanding are such that are to be established in the inner system insofar as they are acquired in the individual occurance, then the question arises how, under the conditions of his chosen method, the experimental psychologist is able to comprehend such inner states at all.
The book, "Verbal Behavior," by B.F. SKINNER (1957), is a classical example of how a purely behavior oriented psychological linguistic theory can look. In a detailed review of this book from 1959 by Noam CHOMSKY, which appeared in the journal "Language", the weaknesses of SKINNER`s approach were, though mercilessly, uncovered. Without going into the details of CHOMSKY`s brilliant analysis, here, it may be mentioned only the core of the critique; CHOMSKY can show that the observable behavior data is, in every respect , too redundant to be able to explain, in a purely behavior based manner, the correlating, intuitive, ascertainable, language relevant processes.
With the means of modern scientific theory, today it would even be possible to begin a rescue opperation of the behavior oriented approach so that it would be at least quasi immune to formal objections. The condition would be that an experimental pschologist would undertake the work to formulate a real theory. C.L. HULL came very close to fulfilling this demand in his main work, "Principles of Behavior" from 1943. The frame of such a theory offers the possiblity to introduce as many of the so-called "theoretic terms" as one wishes. These are theoretic terms that do not directly correspond to emperical givens. But they can still be used freely in the frame of the theory as long as their use does not lead to contradictions within the theory. In the case of non-observable language relevant understanding processes one could make so many methodically correct assumptions hypothetic variables and their interaction with observable data as are necessary to do justice to the speculations about possible internal processes.
This last portrayed action is formally correct, but may be, in the long run, unsatisfactory because its anchoring in phenomena is to weak. Nevertheless, today this way of thinking has its supporters. Under the immpresion of CHOMSKY`s annihilating critique, the paradigm of "cognitive psychology" developed, which was later integrated under the general lable of "cognitive science". In the frame of "cognitive psychology" and "cognitive science" one is ready, beyond the demands of radical behaviorism, to accept again special knowledge relevant processes in observant behavior systems, but in principle one has no more data to use than the radical behavioral researcher. Until now, in the frame of cognitive psychology and cognitive science, no real theories have been developed. The majority of assumptions regarding internal processes would be, by the rule, only partially safeguarded by the creation of a computer program. Such a proceeding can not be convincing. Computer programs are no substitute for complete theories. In the worst case, they are modern versions of fairytales: One constructs with familiar elements something new that corresponds to no reality outside the representation surface.
Another rescue opperation for the insufficiency of a behavior based explaination approach in the case of language relevant data today, is recourse offered through modern physiology, especially neurophysiology. Though the enormous advances in the investigation of the physiological structure of the human being it has become possible to correlate external observable behavior, here called macrostructure, with physiological processes in the body, here called microstructures. Under the prerequisit of the signal processes in the neural pathways and neurons, more complex circuits are reconstructed and the different observable actions are correlated. The elaboration of this correlation is the working field of the "physiological psychology" respectively, "neuropsychology".
The dilemma of behavior oriented psychology would be solved if one was able to prove that it would be possible, through the recourse in physiological processes, to show a functionable substitute for experiential accessable language relevant processes.
What, in any case, is possible, is to make the correlation between acoustic stimuli with typical neural arousel patterns in specific neural pathways and the area of the neural cortex. But, it will be difficult if one wants to neurally classify acoustic stimuli in terms of specific meaning structures. Before one can make a neural classification, one must already have, in this case, a non-neural classification at his disposal. Otherwise, the correlating neural facts in terms of the intuitive accessable linguistic meaning would be uncertain. In other words, the inclusion of physiology, today, could eventually offer the possibility to encode the language internal, experientially given, language relevant processes also in a neural form, but only under the prerequisit of a functioning, language internal, experientially given, world of meaning. If this "language internal world of meaning" is not present, a corresponding interpretation of this neural data is not possible. The inclusion of such a language internal world of meaning would destroy the methodical frame of physiology just the same as physiological psychology. Written statements from test subjects are no substitute for data gained by introspection.
From these considerations, one sees, that the rescue attempts for a behavior oriented approach in the context of language understanding processes failed. Neither the formal increase of theoretic terms to a theory nor the inclusion of modern physiology put the behavior oriented explaination approaches in the position to clarify the interesting processes in the context of linguistic understanding. Beyond the afore mentioned disciplines, this negative result concerns also such popular disciplines as, for example, phonetics, linguistics, computer linguistics, large portions of language philosophy and with that, the total concept of the so-called cognitive science. As SABAH noticed, the pragmatic is also concerned as far as he only deals in a very abstract form with linguistic acts and under intentional conditions, without connecting enough the abstract decriptions to the comprehensible phenamena.
This result is annihilating because it shows that with our usual scientific instuments we are not well enough equiped to reconstuct good enough the linguistic understanding, a central phenomenon for us as human beings. At the same time this negative outcome also gives a plausible reason why the intense research efforts of the last hundred years in linguistic understanding has not moved from its position.
THE IRREDUCIBLE FACT OF CONSCIOUSNESS
Luckily, the failure of the method does not mean the destruction of the phenomenon at the same time. Unscientifically, in our everyday lives, we constantly practice understanding, as humans we can react appropriately to situations, and yes, we can even learn any natural language. And it`s not the case that we carry-out our language activity completely unconsciously. We can bring the relationship between linguistic expression material, (sounds, signs, gestures) and the most different, in part very complex, imaginings, to a certain degree, to a conscious state. We can distinguish different imaginings and dynamics through our imaginations and know that we have them. If we would not have these prescientific given phenomena in the described form at our disposal there would not be any test subject for a possible scientific investigation.
So, if we accept that we have experiential access to language understanding processes and if we, at the same time, accept that a behavior oriented explanation,which does not refer directly to the phenomena, is unsatisfying because it is "arbitrary", then the question arises whether there is a possibility of an explanation of language understanding processes, which approaches at the experiential access of the phenomena that is not completely arbitrary.
The approach of experiantial access leads us to the "method of philosophy". Looking at the variety of philosophies and philosophic methods, I will limit myself here to emphasizing only those characteristics that I find relevant for dealing with language understanding. ( Some important authors that have had an influence on me are: ( R.DESCARTES (1641-47), J.LOCKE (1689), I.KANT (1787), G.W.F.HEGEL (1807), E.HUSSERL (1913)).
(A1) In agreement with John SEARLE, I assume that conscious experiences have there own "ontology", i.e., they only have their uniqueness as subjective experiences. When removed from individual subjective experience the uniqueness is non-existant. (SEARLE 1992, pp. 14-16)
(A2) The amount of experience constitutes the "space of consciousness". Experiences are to the consciousness phenomena and givens (primary givens but not necessarily original givens). In other words, a phenomenon is only phenomenon for the person that "has" it at the moment of the experience. The phenomenon is unquestionable and in this sense primary. In any case, the phenomenon, under the inclusion of "memory", can be identified as a phenomenon in a "stream of experiences". Now, this phenomenon can be classified as "standing in relation to" and through that, as "arranged". As such, it is not necessarily "original" anymore.
(A3) A consciousness is able to know that it has phenomena as phenomena; it is reflexive.
(A4) A philosophy as a conscious method can attempt to explicate from the perspective of knowledge the structure of experiencing with the help of a language. Besides a natural language as a basic language, I suggest the use of a formal language to errect formal structural theories relative to the structure of experiencing. (For example, HINST (1974), (1991-94)).
(A5) Formal philosophic structural theories of consciousness allow themselves to be directly compared through controlled discourse in the basic language with regard to the structural assumptions. The question remains, to what extent, in the case of structural differences, notification of the basic decisions can be achieved (my use of the term structural theory is oriented to BOURBAKI (1970) and LUDWIG (1978)).
(A6) As additional means of control and communication of philosophical structure theories of the consciousness, computer simulations are set-up which are constructed with the handicap of the theory taken into consideration. The object of such a simulation is an intelligent agent that simulates the structure of the human consciousness as far as it is described in the theory. (see below)
(A7) With (A1)-(A6) is intended the model of a " formal computer supported philosophical structure theory of the consciousness" which understands itself not as the opposition to a natural scientific theory formation. By centralizing the theme of the observer from his perspective of self experience, the creation of a fundemental science, in the sense of a generalization of LUDWIG`s fundamental physics ( p. 13 and following) is initiated. That means, the connection between the empirical and conscious oriented science begins where the empirical sciences are ready to think explicitly about epistomological prerequisits of empirical theory formation. If the above assumptions are correct, the result of such explicit thinking allows itself to be adequately articulated with the formal philosophical structure theory of consciousness which is described above. Empirical theories allow themselves then, to be reconstructed as special part theories within such a philosophical theory of the consciousness. The possible parallelism of a physiological theory of the body, including the neural systems, with a pholosophical theory of the consciousness would be a special case here. In a correct way of action it should be possible to project both theories onto eachother without the meaning that one theory is a substitute for the other. The still existing opposition between the liberal arts and the natural sciences would have to vanish in the long term as a methodical artifact.
CONSCIOUSNESS BASED AGENTS
Subsequent to a so-characterized philosophical method and under the prerequisit of the preceeding considerations, I give forth to the following two theses:
(T1) Intelligent agents that shall communicate in a human form, must be consciousness-based agents. The same is valid for robots.
(T2) The use of consciousness-based agents requires the existence of suitable formal philosophic theories of the consciousness.
From (T2) it follows that in the future the traditional KI will no longer be the primary discipline for the use of intelligent agents, but a philosophy that understands itself as a basic science for consciousness-referred model formations. Just as mathematics and logic are, computer science will be, a possible tool of philosophy.
THE KNOWBOTIC INTERFACE PROJECT
There will possibly be different ways to put this so-outlined research program into practice. At the Institute of New Media since the summer of 1994 we have been persueing the following double strategy called " Knowbotic Interface Project (KIP)" :
(S1) We are attempting to develop a formal philosophical structure theory of consciousness.
(S2) Parallel to that we are driving forward the development of a suitable computer simulation.
What the last point, computer simulation, concerns, is that it was possible to complete the prototype of Knowbotic Interface, Version 1.0. It can be tested in the Institute of New Media. From February,1996 on, access to testing shall be at public disposal over Internet.
The version 1.0 prototype is not connected with the demand to simulate a philosophical theory of consciousness. Through-out the development it was very important to us to create a sort of software-construction-kit. The construction-kit allows us to put all the characteristics of the formal theory of consciousness that the theory demands directly in the program in the next phase of development. Phase two of the Knowbotic Interface Project is planned to begin in December 1995 and will run until November 1997.
In spite of its temporary character, the prototype already offers all important basic elements that will also continue to be valid in phase two.
The following components are the most important:
(E1) First, naturally there are programs that are supposed to simulate consciousness structures. We call such a program a "knowbot (=to know+robot)". In phase one a knowbot is realized as a reactive system, that means, still without reflections. It has
- sensory channels (hearing, smelling, tasting, touching, not seeing!), it has
- bodily conditions or needs/emotions (hunger, thurst, tiredness, fear, aggression, sex drive, curiosity), it has
- a simple memory that forms simple object-hypotheses based on sensory impressions including the other "consciousness conditions" and organizes these in the form of "road maps", it has
- a simple need-driven activity component that organizes its elementary actions (going, turning, taking, putting in the mouth, eating, drinking, sleeping, mating, letting go of objects) in action sequences corresponding to its actual needs or emotions and it has
- a simple language module for simple one word sentences. Without a coach a knowbot builds-up -relative to its own world experience- its own "vocabulary" that can be used by a pseudo-knowbot to communicate with the knowbot itself. But a coach would be able to "induce" the vocabulary of a certain language (German, English, Arabic, Russian, Chinese,...) into the knowbot by constant persuasion; a knowbot can learn more than one language at the same time.
The idea to train a learning capable computer program by letting it go through the world like a human child and being raised by human beings was already formulated by TURING (1950, pp,177-179). But he didn`t have the adequate technical resources at his disposal to put these ideas into functional experiments.
(E2) In order for knowbots to learn any language like children and then to be able to use them, they need surroundings that on one hand, simulates sensory stimuli and on the other hand, assimulates actions. Thereby, the sensory stimuli and also the assimulation of actions should happen in a way that as close as possible corresponds to the world experience of human beings. A program that achieves this we call a "world process" or simply a "world". In phase one we experimented with worlds that portrayed islands in the ocean furnished with trees, bushes and plants. There was only one simple kind of animal, a preditor, that attacked the knowbots. Every object in the world has a so-called "sensory shell", i.e., every object, besides its geometrical form, is also characterized by taste, smell, touch-values and sound-values.
(E3) The third element is the "pseudo-knowbots". A pseudo-knowbot is a program that is started by a human user. With this program one has the opportunity to register in the world- process like a knowbot and then, to live in the simulated world like a knowbot. From the view of a knowbot a pseudo-knowbot appears like a knowbot. The only difference is that in the inside of a pseudo-knowbot, instead of an artificial consciousness a human consciuosness is at work. The first pseudo-knowbot called Stevie I, run under LINUX, has the disadvantage that it doesn`t allow a view inside the world; the user finds himself in a quasi "blind flight", he has only his diverse sensory data at his disposal to orient himself. Stevie II, available starting in January, 1996, runs under MS Windows, and also offers a view in the actual surrounding of the respective user. With that, a user is put in the position to initiate contact with a knowbot in a world from home, and to interact with the knowbot bodily and linguisticly.
Besides these main programs, there is a line of additional help programs, for example, an independant "WorldView" and a "WorldEditor" (the afore mentioned program available starting in March 1996).
All the programs are, moreover, network capable, i.e., every named program can be started anywhere in the internet: a world-process in Frankfurt, a knowbot in London, a pseudo-knowbot in New York and a world-view in Sidney with simultanious emitions of the world- view over a television station.
The experiences we were able to collect during the contruction of version 1.0 of the prototype, were very important to plan the subsequent proceedure.
OUTLOOK ON PHASE TWO
Finally, I`d like to take up three topics that are of special interest here.
The first topic concerns the world process. Here, the question is asked whether the world process should be realized locally or spread-out, and how far an achievable mathematical model is available that easily and quickly allows it to callculate the collision points of many complex, self-moving objects under consideration of individual points on their surface. The problem of collision calculation is to be solved in local and spread-out worlds. Until now, we`ve only solved the problem of the cylinders. Eventhough the capacity of local worlds, through the parceling of the worlds relative to an amount of processors (i.e., on multi-processor machines) , could be easily increased, it would be very desirable, if it could be cleareded-up whether the world process could be realized as a spread-out process, at least as far as collision calculations are concerned.
The second topic concerns the language capablities of the knowbots. In contrast to the preceeding considerations, in which we concentrated on the problem of language-understanding and language-producing, we have established the additional requirement for the knowbot that their language competence must be acquired independantly from point zero. From knowbots we require that they should be able to "learn any natural language just like children": We see the "understanding" of language just like the "situation refering production" of language, as a partial aspect of this comprehensive ability. With this in mind, we assume from the working hypothesis that a language can develop itself ony relative to a world knowledge that was or is simultaniously in existence at the time of that development. This world knowledge, to a certain degree, is independant from language. If, for example, children begin to produce one-and then more-than-one word sentences then, they are not blank slates; to the contrary. They very obviously already have an initial structured world experience to which they begin to build-up their new language structure ( in this sense, for example, KEGEL (1987), OERTER/MONTADA (1987)).
One can already clearly see in the prototype this dependancy of the language structures on the available differentiation in the area of sensory data, memory, bodily conditions, emotions and plan of action. Eventhough the prototype knowbots from phase one are able to form various one-word sentences or accept them from others and they also can articulate them appropriate to the situations, they are still at this time completely unable to learn real more-word sentences and use them. The reason for that is that right now, they don`t have any suitable mechanisms at their disposal to be able to make the internal differentiation of their world experience, which is necessary to form more than one-word sentences. Such a differentiation requires a complex cooperation between memory structures, situation representation, plan of action and part of language. But the development of the prototype, put us in the position to be able to say, pretty exactly, how these missing differentiations must be procurred.
The third topic touches the social dimension of our theory formation. In October, 1995 , it was decided that we make the Knowbotic Interface Project into an internet project. Besides the more practicle considerations that , of course, we could increase the development capacity for the project, the basic considerations stand in the forground that the "philosophical project of the self-description of the consciousness " actually concerns everybody; if people, on the one hand, continuously build-up technological possibilities, so far that they are already in the position to be able, not to build-up new worlds but to destroy them, then humans should force themselves to come to a possibly complete knowledge about themselves, and that in a way that finds a possibly complete acknowledgment and with it communicative obligingness.
As our preceeding considerations have shown, empirical research is not sufficient enough for this; we also need an analysis of the human cognition process under the conditions of this realization. How could that better happen than through a world wide internet discussion, which is open to everyone, that results in disscusable formal theories, and beyond that, would be also vivid through computer simulations, That way, even someone who doesn`t understand the formal theory is in principle in the position to criticize the presented model on the basis of his personal world experience.
How far this demanding project in the end will be feasible, at this time, nobody exactly knows. The only thing that is clear is that even only a partial success of this project would mean on one hand, a decisive step forward in the direction of "understanding the understanding" and on the other hand, in the direction of a new form of "collective intelligence", in which real and artificial subjects interact in a way that we haven`t even seen in science fiction.
Bibliography
Comments are welcomed to kip-ml@inm.de
Daimlerstrasse 32, 60314 Frankfurt am Main, Deutschland. Tel +49- (0)69-941963-0, Tel-Gerd: +49- (0)69-941963-10