Abstract

I introduce the notion of knowledge relativity as a proposed conceptual link between different scientific disciplines. Examples from Informatics and Philosophy, particularly Newell's knowledge level hypothesis and Popper's world 3 of knowledge, are used to demonstrate the motivation for making this notion explicit.

-

Not every assumption is knowledge and whether somebody’s personal views on the world are classified as knowledge is determined by complicated interactions among human subjects and between human subjects and their environment.

The word science itself means knowledge and thus one would assume that the distinction between arbitrary unilateral assumptions on the one hand and shared mutually assured knowledge on the other hand is uncontested in contemporary science. But one scientific discipline has adopted a notion of knowledge that abandons the distinction between knowledge itself on the one hand and the mode of it’s expression or temporary attempts to arrive at it on the other hand: In Informatics, knowledge representation is progressively equated with knowledge itself, and isolated pieces of knowledge representation stored in a single computer system with mutually confirmed truth. Philosophy would be ideally situated to contribute to a more mature notion of knowledge in Informatics, but the relativity of knowledge appears to be too commonplace in contemporary Philosophy to be explicitly stated.

In this article, I will therefore try to explicate knowledge relativity and the potential role of this notion both in Informatics and in Philosophy. The starting point for this task will be asking the missing ”Who” question for some of the basic concepts in Informatics and the result will be a notion of knowledge that emphasizes subjects over objects and involves an understanding of standardization instead of the predominant notion of truth. Since it’s early beginnings, Informatics as a scientific field has favored concepts and models that allow for the marginalization of subjectivity and relativistic views. One of the numerous examples for this tendency is Newell’s hypothesis of a knowledge level as the sole and exhaustive location for knowledge in a system with artificial intelligence (Newell 1982) . This level would have all the properties of a normal computer system level: among other properties, it can be implemented without references to internal details of other levels, and it can be reduced to the level below it (the symbol level) by defining it’s medium (knowledge), components and laws via those (symbols) of the level below it. According to this hypothesis, knowledge could be seen as an abstract property of (computer system or human) agents implementable via various different forms of symbolic representations in the same way that symbolic representations are implementable via various different forms of electronic hardware. Such an understanding of knowledge would make it impossible to directly verify or falsify the presence of a specific element of knowledge in an artificial agent: the highest directly observable system level contains symbols that might or might not encode a specific knowledge, but the knowledge itself would reside one level above those representations and would be removed from direct observation. But Newell proposed a mechanism of indirect verification for the knowledge level: If some (human or artificial) agent A can detect the impact of some specific knowledge in the actions of another agent B, agent A can verify the presence of such knowledge in agent B (Fig. 1). Through this hypothesis, Newell

Knowledge K'

Symbol

System S'

Actions

Goals G

Knowledge K

Symbol

System S

attempted to end an internal dispute among artificial intelligence researchers, a controversy centered around the question of the ”best” form of knowledge representation. According to Newell himself (Newell 1993) , this attempt was partially successful, and a significant part of artificial intelligence research has been (implicitly or explicitly) based on the knowledge level hypothesis. I will not discuss the correctness of Newell’s hypothesis in this article, partially because it is too complex to determine how it could be verified or falsified at all, but also because the correctness or incorrectness of this hypothesis is not directly linked to the central argument in this article: the general tendency of Informatics to suppress notions of relativity and the need to introduce explicit terminology for expressing relativity. A starting point for detecting the implicit objectivist world view (Stamper 1993) underlying artificial intelligence concepts like the knowledge level is to focus on those aspects that authors like Newell do not discuss, and the questions that are not answered. A good example is: ”Who is the agent that will verify the presence of knowledge in another agent?” More than one agent could take the role of the observing agent A in Newell’s knowledge verification mechanism and the different As could come to different conclusions on whether knowledge is present in agent B (Fig. 2). Newell does not explain how the A agents would be able to come to any consistent conclusion on agent B’s knowledge, and since their symbol levels might be mutually incompatible, it is unclear how they could even communicate their different views on agent B. The only conceivable way Newell’s knowledge verification might yield consistent results is by relying on a standardized observer agent A that serves as the absolute reference on detecting knowledge. Newell implicitly assumes objective knowledge that would be the basis, rather than the result of attempts to create artificial intelligence. But if such a standard agent existed, Newell’s entire knowledge verification procedure might be obsolete before it is ever applied: this standard verification agent would already incorporate the perfect embodiment of intelligence and knowledge, and any attempts to build more intelligent agents would have to fail. The main effect of Newell’s knowledge level hypothesis is not what it accomplishes, but what it prevents: the authority to interpret symbolic knowledge representations is restricted to computer systems, since their (virtual) behavior is the only way for determining what knowledge is represented. Human subjective interpretation is excluded from the process of symbol interpretation, since Newell’s artificial agents never directly expose their symbol level to human agents. Newell’s approach of replacing human subjective interpretation with the implicit assumption of an objectively knowledgable observer represents the general trend in Informatics, and both method and motivation of this approach can be traced back to the origins of the discipline: The original technical definition of information (Shannon 1948) is based on the ability to predict the next symbol in a stream of communication. Again, the open question is: ”Who will predict the next symbol?” Different potential receivers of the same symbolic message will have different knowledge about the world in general, the language used for sending the message, and the sender. Thus different receivers will have different abilities to predict the next symbol, and the information content of the same message will potentially be different for each receiver (Fig. 3). Shannon arrives at an objectivist result by assuming the existence of a subject-independent dictionary containing absolute probabilities for a given language. The receiver of Shannon’s mes

Symbols S

Symbols S'

Channel

Information

Sender

Receiver

sages turns out to be the same implicitly standardized perfectly knowledgable observer that Newell uses to verify the existence of knowledge in computer systems. Shannon, like Newell, was motivated by the goal of eliminating subjective interpretation from his model of symbolic message content, and his information notion is still in use today. Informatics has suppressed knowledge relativity since it’s beginning, and current trends to equate knowledge with knowledge representation are the direct continuation of this tradition. The discipline initially had good reasons for choosing this approach: At the time of Shannon, technical reliability of communication was the primary goal, and side effects on the potential processing of stored knowledge inside computer systems were largely irrelevant to the engineers that founded Informatics. Wherever issues of knowledge relativity come in conflict with engineering properties like reliability, predictability, and consistency, Informatics has generally given preference to those views that favor engineering goals. Implicitly, knowledge relativity has always played a role in Informatics: as the notion that has to be eliminated.

Three Worlds

Due to it’s longer history, Philosophy has witnessed more changes in the role of knowledge relativity than Informatics. During it’s co-existence with Informatics in the last 60 years, however, knowledge relativity played a comparatively implicit role in Philosophy, much like in Informatics. But in contrast to Informatics, knowledge relativity was implicitly treated as a given basis of inquiry. To illustrate main differences in the implicit

World 3

World 1

World 2-1

World 2-2

treatment of knowledge relativity, I will discuss a prominent example from the Philosophy of Science: Popper’s 3 worlds thesis (Popper 1972) . Popper defines 3 different worlds that are interrelated but separate (Fig. 4): World 1 is the (observer-independent) physical universe, world 2 is an observer’s image of world 1, and world 3 is the collective symbolic representation of world 2 shared among a multitude of observers. World 3 is necessarily embedded in world 1, since observers are only able to communicate each other’s symbolic representations if they are able to physically perceive them. Furthermore, any action performed by one of the observers can be interpreted as a contribution to world 3, and scientific experiments are the most prominent example of this feature. Popper’s motivation for his 3 worlds thesis was somewhat similar to Newell’s motivation for his knowledge level hypothesis: Popper wanted to show a path towards objective observer-independent knowledge. But in contrast to predominant Informatics approaches, Popper assumed subject-relative knowledge and subject-relative representation as the starting point of this path. Objective knowledge is only achieved as the result of a long process of negotiation among subjects. This process is permanent, since world 3 only approximates world 1, without ever reaching total consistency with it. 5

A Matter of Perspective

The reception of Philosophical concepts in Informatics is somewhat ambiguous: Informatics approaches frequently borrow their terminology - for instance ontology (Gruber 1991) - from philosophy, but later transform the meaning of these terms in very drastic ways, at times even reversing their original meaning. I argue that this phenomenon is meaning 1

meaning 2

Signs

Human

Machine

due to the dual role of symbolic representations in Informatics: During computer system programming, symbolic representations are created and interpreted by humans, but during computer system execution time, these representations are interpreted and modified by machines. The algorithmic sign (Nake and Grabowski 2001) (Fig. 5) incorporates a dual role towards the human observer on the one hand and the computer system on the other hand, at times as an external symbol used by human agent(s), and at times as an internal symbol used by machine(s). Machines are standardized products, and they are expected to work according to specifications. The same symbolic representation would therefore be expected to be open for re-interpretation when read and analyzed by different human agents, but would be expected to be stable and rigid in it’s meaning and function when processed by different computer system agents. Since the final goal of Informatics is creating computer systems, the discipline has typically favored views that marginalize human subject-relative aspects. This approach has proven successful in areas that require predictability and consistency, but unsuccessful when the cost of implicit standardization was too high. One example for the latter are knowledge management tasks or the computer support of human creativity. For a more successful interaction with computer systems in these areas, knowledge relativity has to be made explicit. 6

Dimensions of Knowledge Relativity

In order to explicate the relativity of knowledge I will try to list the different dimensions of knowledge relativity: 1. subject who is holding the knowledge: who knows? 2. representation what kind of symbolic expression is used to communicate the knowledge: what was expressed? 3. evaluator who is judging the presence of knowledge: who thinks someone knows? 4. communicative intent was the knowledge expressed in order to inform or in order to reach agreement on known issues: do we need to discuss this? 5. functional intent what use was intended for the representation used: what will it change? 6. receiver who was the intended receiver for the representation used: who should know? Not all of these dimensions need to be fully developed. A human agent might have knowledge without communicating it in any form, for instance, so the representation dimension would not be developed. Some dimensions might also have identical values, the subject holding the knowledge and the evaluator judging it’s presence might for instance be identical (”I know”). The primary line of distinction in the treatment of knowledge between human communication and machine intelligence can be found in the functional intent (will the representation be interpreted as an external signal or directly processed inside the system?), and the primary line of distinction between different disciplines like Philosophy and Informatics can be found in the communicative intent (do we want to declare a standard or start a discussion?). Such a definition of

Knowledge K1

Knowledge K2 t n e t n i l a n o i t c n u f

Knowledge K3 t n e t n i l a n o i t c n u f

Information I1

Information I2

Representation

Sender

Receiver R1

Receiver R2

knowledge relativity does not require reference to any object, but it requires reference to subjects, as can be demonstrated on the example of a knowledge relativity aware version of Shannon’s information (Fig. 6): A signal (the representation is assumed to already contain any channel distortion, S’ in Fig. 3) sent by one subject could be received by a multitude of subjects. For each receiver, this signal could represent different knowledge, always in relation to each receiver’s knowledge about the sender and each receiver’s intentions for this signal. Thus each receiver could detect a different amount of information for the same signal, and the same signal would contribute to the knowledge of different receivers in different ways. For an individual receiver, Shannon’s information definition as the degree of un-ability for predicting the next symbol (in relation to this receiver’s knowledge) still holds. From a knowledge relativity aware perspective, knowledge is not contained in any of the involved subjects, nor can it be assembled from or reduced to subject-external components like those constituting symbolic representations: Since any understanding of the link between some specific subject, some specific representation and some specific knowledge always requires some other subject in the role of evaluator, the link between representation and knowledge will always be dynamic. 7

Conclusion

The main motivation for the introduction of knowledge relativity in this article was to facilitate the dialogue across disciplines, particularly between Informatics and Philosophy. But assuming a fertile dialogue and assuming this new notion proves useful, I would expect a direct effect both in Informatics and in Philosophy, and potentially in more disciplines. A fair amount of literature outside Informatics and Philosophy deals with issues similar to the ones exemplified above, and if it is true that we already live in a ”knowledge society”, some additional clarity on the nature of knowledge or at least the nature of the term knowledge should prove useful.

References

Gruber , T. R. : 1991 , The Role of Common Ontology in Achieving Sharable, Reusable Knowledge Bases, Principles of Knowledge Representation and Reasoning , Morgan Kaufmann, Cambridge, USA.

Nake , F. and Grabowski , S.: 2001 , Human-computer interaction viewed as pseudocommunication , Knowledge Based Systems 14.

Newell , A. : 1982 , The Knowledge Level , Artificial Intelligence 18 , 87 - 127 .

Newell , A. : 1993 , Reflections on the knowledge level , Artificial Intelligence 59 , 31 - 38 .

Popper , K. : 1972 , Objective Knowledge : An Evolutionary Approach , Oxford University Press, UK.

Shannon , C. E. : 1948 , A mathematical theory of communication , Bell System Technical Journal 27.

Stamper , R. K. : 1993 , Social norms in requirements analysis , in M. Jirotka, J. Goguen and M. Bickerton (eds), Requirements Engineering, Technical and Social Aspects, Academic Press, New York, USA.