Knowledge Representation in Artificial Intelligence

Preface: This paper was written in the fall of 2019 for my Epistemology class at Lewis & Clark College. The course focused mainly on the theory of knowledge, and I wanted to incorporate my AI ethics research into my term paper — here is the result.

Section 1: Introduction

Over the last century, rapid advances in computing have raised important philosophical questions concerning the nature, abilities, and roles of technology in a modern human context. Given that most of the advancements of computing machinery stem from its information processing abilities, epistemic questions are among the first to arise. We have understood and gained so much from their powers, but still have so much to learn — most importantly, that these machines cannot learn in the same way humans do.

I will outline a basic account of what knowledge is, and provide a fundamental explanation of artificial intelligence and the challenges of computing machinery. I advance the position that what we often misdiagnose as knowledge is merely information processing, and a great deal of the literature on knowledge representation is guilty of this anthropomorphic conflation. At best, artificial intelligence offers us a functional equivalent of knowledge, though it can never truly be knowledge. Our abilities do not, and should not, match each other. However, there are important analogous abilities shared between artificial intelligence and human knowers, and externalist accounts of knowledge coupled with information theory is invaluable in our understanding of what these abilities can help us achieve.

Section 2: Foundations of Artificial Intelligence and its conception by humans

Definitions of artificial intelligence often vary from context to context. It is usually defined in the negative as intelligence that is exhibited by a machine or other artificial entity, rather than something natural like a human. It is usually further defined in the context of humans as attempting to emulate human cognitive abilities in computing systems, mimicking things like reasoning, learning, problem-solving, and decision-making through data processing and pattern analysis.

Artificial intelligence usually goes hand-in-hand with machine learning, which is a machine’s ability to change its own instructions in the pursuit of a given goal. This is done using algorithms, which are basically sets of rules for the computing system to follow. It takes in inputs in the form of raw data, and processes them based on these rules to achieve the desired outputs. This creates the effect that the computer is learning, as it alters its treatment of the inputs to come as close as possible to the desired output representing its goal.

AI has grown into a helpful and widely used tool in the advancement of technology. In developed countries, daily interaction with AI is not uncommon, from personalized recommendation systems such as Spotify, financial dealings such as applying for a loan, GPS navigation, and medical diagnostic systems, to name just a few. These dealings can be difficult to notice due to a problem called the AI effect: as soon as an AI is capable of doing something, it is no longer considered AI. Once the performance exists, it is seen more as a mere computation or function than something we would characterize as intelligent.

This becomes even clearer when put into the context of media’s portrayal of AI as a science fiction-like entity, a wolf in sheep’s clothing ready to enslave humanity as soon as it can gain power over us. This conception is misguided in a number of ways, but most importantly it fails to identify the type of AI we are capable of making.

AI is separated into two categories, which go by various names. The first is narrow or weak AI, which is AI that can do one or a few things very well, but not much else besides that. General or strong AI is AI that can perform many things well, almost like a human can, as we are generally intelligent beings. As of our current technology, general AI does not exist and will not for a long time. AI also has no evolutionary reason to act out of self-preservation like an evolved biological being would. On top of all this, humans are notoriously bad at predicting the rate of development of AI. Our fears appear to stem from underestimating what we’ve created, but it could be equally as dangerous to overestimate the abilities of artificial intelligence as it is to underestimate them.

From these descriptions, it becomes apparent that there is a semantic confundity going on here. Using terms like “intelligence” and “learning” to describe computational processes implies more ability than is actually there. We lack a proper vocabulary to describe what’s occurring, so we’ve co-opted our vocabulary for describing our own cognitive processes to try to capture what we’ve created. Neural networks, artificial intelligence, machine learning — all guilty of abusing cognitive terminology to describe a similar process that is too different to warrant the same word use.

This semantic confundity is a symptom of a larger cause: We have a hard time understanding what this technology is and what it’s capable of. That symptom is furthering our conceptual misunderstanding by failing to capture what is there. This is of great epistemic concern: AI directly influences our epistemic processes, and has important parallels with them. But if we fail to capture what is truly going on with AI, we will continue to confound ourselves, hold false beliefs about one of the most important technologies of the modern age, and miss out on important opportunities to learn more about ourselves as informationally driven beings.

To demonstrate how this confundity is directly connected to epistemic concerns, we turn to the literature on the theory of knowledge and knowledge representation.

Section 3: Knowledge and knowledge representation as a case of confundity

Our traditional understanding of what knowledge is consists of three components: truth, belief, and warrant. For our purposes, we will pay little attention to truth and treat it as brute. Belief is a product of cognitive behavior, and cognition is mental representations of how the world is. Warrant is how we offer a defense of what we believe and is true, securing the final connection needed for knowledge. The accounts of this are many, but an important one is internalism versus externalism. Externalist accounts posit that one is justified in their belief by their relations to the world. This does not offer justification, but really just warrant. We will see that for our purposes, this is a good thing.

Knowledge representation is the modeling of knowledge in a way that does not qualify as knowledge. Human knowers usually express knowledge propositionally, but knowledge representation does not have to be formatted as such. In the knowledge representation literature, it is often depicted with types of formal logic systems.

What does this have to do with artificial intelligence? Not only do the knowledge representation literature and discussions of AI commit the same conflation of cognitive terms to non-cognitive entities, they sometimes do so in junction, as knowledge representation is used in our conceptions of AI. Through this, we can examine knowledge in the context of AI.

First, a clarification. AI does not have emotions, cognition, sentience, consciousness, or the ability to hold any kind of affective state. Therefore, as of our current technological abilities, AI cannot be knowers because they cannot satisfy the belief condition of knowledge. This is why we are able to address knowledge representation in AI, but not knowledge itself.

That being said, the literature on knowledge representation is guilty of using the same confundity with AI in the form of knowledge attributions, which entails belief attribution. Difficulty arises when we consider the term “knowledge representation” — where are the components necessary for knowledge? How is truth, belief, and warrant all connected to it? Under a proper theory of knowledge, in order for this to qualify as knowledge representation, it would have to have some sort of account as to how it is connected to the components that yield knowledge. Without these components, it cannot be claimed as knowledge.

In “Sophisticated Knowledge Representation and Reasoning Requires Philosophy,” Selmer Bringsjord et al. show the confundity outright: “Knowledge-based systems (KBSs), then, can be viewed as computational systems whose actions through time are a function of what they know and believe.” In the abstract, they make it clear that the computational systems that this excerpt mentions is referring to either a human or a computer.

One of the most prominent examples of this confundity occurs in Nicola Lacey and Mark Lee’s “The Influence of Epistemology on the Design of Artificial Agents.” Lee and Lacey encoded four different artificially intelligent robots according to the theoretical structures of four epistemological theories, and tested them with experiments to see which ones performed the best in their treatment of information. This experiment is extremely important and enlightening, and we will use their conclusions to our advantage later on, but Lee and Lacey are also guilty of anthropomorphizing. The AIs in question are attributed as having beliefs and knowing throughout the paper.

Section 4: A Solution to the Confundity Problem

There is a simple answer to all of this confusion: recognize that what we are after is being expressed as information, not knowledge. Does a thermometer know what the temperature is? No, but it certainly conveys that information. This example is not to equate the computational abilities of AI with a mere instrument, but to show that a lack of sentience indicates a lack of knowledge, no matter the complexity of the thing in question.

Bringing this back to the literature on knowledge representation, this criticism should not undermine the important work that goes on in the area. It ought to simply be called what it is, which is information representation. This is by no means a simple task, nor does it indicate any sort of loss of meaning by not being referred to as knowledge. It is more accurate, and can aid in our understanding of how information is represented across systems. Most importantly, and for the purposes of this paper, it does not advance an inaccurate account of what knowledge is, and will thus aid in our understanding of knowledge through this clarification.

There is good news on this front. Fred Dretske’s account of knowledge and the flow of information contains important features that can help us fully account for the treatment of information in an intelligent system, be it human or computer.

Dretske’s aim is to develop a philosophically useful theory of information that shows the intuitive importance of information both in philosophy and cognitive science, as well as to bridge the two disciplines. Dretske begins by explicating Shannon’s mathematical theory of communication to give us a useful theory of information. He then applies the theory of information to various philosophical problems: knowledge, skepticism, and perception. Lastly, an information-theoretic analysis of propositional attitudes, which is where he covers beliefs, saying that the instantiation of neural structures gain an information-carrying role through learning.

Some epistemologists hold that Dretske’s account of belief is insufficient to make this a complete theory of knowledge. For our purposes, this account works in our favor. The things he refers to in the belief category have functional equivalent analogs in AI, so we don’t need to worry about his use of belief in a theory describing things without belief. Dretske characterizes belief with information processing ability by the instantiation of neural structures, as noted above. We can already emulate neural structures in AI through the use of various types of neural networks, and we have seen that they can “learn” in the way they process information. This creates a functional equivalent of belief under Dretske’s definition, which secures the most difficult component in creating a clear parallel between AI and natural intelligence.

Certain epistemologists also take issue with Dretske’s denial of closure. If we dip again into the knowledge representation literature, we will find that this also works in Dretske’s favor when applying his theory to AI. In “Knowledge Representation: Two Kinds of Emergence,” Veikko Rantala describes a facet of training a neural network:

“However, the fact that a network learns to recognize that something is the case does not, of course, imply that it would recognize that something else is not the case, that is, a positive recognition does not imply the corresponding negative one. Sometimes it is important to learn both; it is important to learn to distinguish between cats and non-cats.”

This offers an important parallel to our analysis. Recognizing something without recognizing its entailment, in this case its negation, arises when closure is denied. Ian Evans and Nicholas Smith make it clear that this does not bode well for a theory of knowledge. A problem that arises in a Gettier case of Fake Barn Country II is that Henry knows he sees a blue barn without knowing he sees a barn; he knows the conjunction, but not its singular. It is sufficiently similar to conceive of something’s negation to be part of an entailment that would usually create a problem in the denial of closure, but based on our understanding of AI and Rantala’s helpful analysis, this works in nearly the exact way that AI processes information. We gain yet another helpful parallel for understanding AI’s information-processing abilities through Dretske’s account, even though that condition may fail in a human application.

Dretske’s theory most importantly accounts for coherence within a background system:

“To design a pattern-recognition routine for a digital computer, for example, is to design a routine in which information inessential to s’s being an instance of letter A…is systematically discarded (treated as noise) in the production of some single type of internal structure, which, in turn, will produce some identificatory output label (Uhr 1973).”

Dretske’s k represents a background system of information to which new information must cohere. The systematic discarding of inessential information as noise is a way of achieving local coherence within an agent’s background system k. This translates well to AI if we look back to Lee and Lacey’s experiments: The weakly holistic agent performed better than the strongly holistic one because it was better able to discard irrelevant information as noise. The notion of local coherence is well-supported in epistemology literature. In addition to this, most knowledge representation literature relies on the concept of a knowledge base, usually expressed as KB, which can serve in the exact same way as Dretske’s k. This secures our final connection to the relevance of Dretske’s theory in the treatment of AI.

Section 5: Concluding Remarks

We have seen that the conception of artificial intelligence is an important one in our treatment of information. The increase in the informational nature of the world, the prevalence of data in our daily lives, and knowledge representation in our digitized interactions will only serve to increase the importance of understanding epistemology and what it means to have knowledge.

____________________________________________________________

Works Cited

Afrouzi, Amin Ebrahimi. “The Dawn of AI Philosophy.” Blog of the APA. American Philosophical Association. 2018. Accessed 17 December 2019.

Bringsjord, Selmer, Micah Clark and Joshua Taylor. “Sophisticated Knowledge Representation and Reasoning Requires Philosophy.” Philosophy’s Relevance in Information Science. 2009.

Dickson, Ben. “What is Narrow, General and Super Intelligence.” Demystifying AI. TechTalks, 2017. URL: https://bdtechtalks.com/2017/05/12/what-is-narrow-general-and-super-artificial-intelligence/

Dretske, Fred. “Precis of Knowledge and the Flow of Information.” Behavioral and Brain Sciences 6 (1):55–90. 1983.

Evans, Ian, and Nicholas D. Smith. “Knowledge”. Polity Press , 2012.

Lacey, Nicola J. and Mark H. Lee. “The Influence of Epistemology on the Design of Artificial Agents.” Minds and Machines 13: 367–395, 2003.

Lacey, Nicola J. and Mark H. Lee. “The Epistemological Foundations of Artificial Agents”. Minds and Machines 13: 339–365, 2003

Rantala, Veikko. “Knowledge Representation: Two Kinds of Emergence.” Synthese 129: 195–209, 2001.

Smith, Nicholas D. PHIL-311: Epistemology. Lewis & Clark College, Fall 2019.

Wikipedia entry on Artificial Intelligence: https://en.wikipedia.org/wiki/Artificial_intelligence

Wikipedia entry on The AI effect: https://en.wikipedia.org/wiki/AI_effect

Wikipedia entry on AI Winter: https://en.wikipedia.org/wiki/AI_winter