Does Artificial Intelligence Have Concept? ^†

Abstract

“Concept” is a commonly used word, but there are different opinions on what a “concept” is. There are those who study “concepts” from the field of cognitive science and those who study “concepts” from the field of philosophy. In recent years, people in the field of artificial intelligence have also begun to study “concept learning” related to “concept”. What is “concept learning” in artificial intelligence? Can artificial intelligence really have the same “concept” that humans have? Around these questions, this paper will analyze “concept learning of artificial intelligence” and “concept” respectively, so as to preliminarily answer the question “does artificial intelligence have concept”.

1. Concept Learning in the Field of Artificial Intelligence

The artificial intelligence (AI) in this paper refers to the machine ability that simulates human cognition.

In order to become an artificial moral agency and be able to determine morally correct behavior in a given situation, AI needs to have a certain understanding of the concepts related to morality [1]. In his paper “Concept Learning for Safe Autonomous AI”, Kaj Sotala attempted to implant moral concepts such as “rights” and “well-being” into autonomous artificial intelligence. Its strategy was to study the mechanism of human cognitive concept and then imitate human cognitive mechanisms in the field of artificial intelligence [2].

Kaj Sotala introduced different theories about the study of human concepts. For example, one theory studies the concept from the perspective of grammar. According to this understanding, one can equate the process of learning the concept of “effective English sentences” with learning grammar. Some theories are based on the general theory of representation, which argues that concepts can be represented as geometric structures in multidimensional space. In a word, by formalizing and structuring concepts, the mechanism of concept learning is found and can be applied to artificial intelligence. The above method is the feasible framework for AI to learn moral concepts, but Kaj Sotala does not specify the technical details for how to learn it.

Brenden M. Lake started the first Bayesian program learning in his paper “human level concept learning through probabilistic program induction”. In Brenden’s view, the biggest difference between machine learning and human learning is that machines learning concepts require a lot of data and examples, whereas humans learning concepts require only a small amount of data or a single example. In addition, human beings can generate additional functions through concept learning, such as creating new templates and new categories using existing categories, which machine learning cannot do. Therefore, Brenden wants to mimic human learning and apply it to artificial intelligence. The core problem to be solved is how to summarize a rich concept from sparse data and generate rich representation. The Bayesian program learning (BPL) framework allows a large number of visual concepts to be learned from a single example and generalized in ways that are indistinguishable from human beings.

In BPL, “Concepts are represented as simple probabilistic programs—that is, probabilistic generative models expressed as structured procedures in an abstract description language.” [3]. Brenden used an existing database of handwritten characters to develop a program that could write at one shot and it passed the visual Turing test, making it difficult to tell what was written by the program and what was written by a human. This attempt makes it possible for artificial intelligence to learn the concepts of words.

The AI mentioned in these two articles have the following in common: both of them have mechanisms that are achievable. The AI in the first paper has a complex moral mechanism, and this paper provides a variety of possible methods to search for such a complex mechanism. The AI in the second paper has a simple mechanism, which does not involve morality and emotion and this paper has found and realized this mechanism by using Bayesian program learning.

2. Different Understandings of Concepts

Although people often use the term of concept learning in artificial intelligence, it is still not clear what “concept” is. What is “concept”? This question can be answered from three perspectives: common sense, cognitive science and philosophy.

2.1. From the Perspective of Common Sense

From the perspective of common sense, “a concept is merely a symbol, a representation of the abstraction.” [4]. Any representation or symbol can be called a concept. This is the most general understanding towards concept. Since AI are based on symbols or representation, in this general sense, AI have concept.

2.2. From the Perspective of Cognitive Science

In cognitive science, there are three main concepts, that are prototypes, exemplars, and theories [5] (pp. 76–77). In Edouard Machery’s view, these three concepts are three entities, which have nothing in common and they are three distinct concepts. These three entities have been shown to exist in the brain and are often used in the different cognitive processes. Since the attempts to enforce the explicit definition to concept have already failed, concept is something similar to the heterogeneity hypothesis and is not a natural kind. This paper will explain these three concepts respectively and then show whether AI has these three concepts.

The prototype theory is that concept is prototype, and the prototype is “a body of statistical knowledge about the properties deemed to be possessed by the members of this class.” [5] (pp. 83–84). Prototype theories of concepts vary depending on how they describe the nature of the statistical knowledge stored in the prototype. The prototype can not only represent the typical properties of categories, but also can represent the cue-valid properties of categories. According to prototypes, people can judge whether one thing belongs to one category by checking the similarity between the prototypes and the thing. The prototype theory is very similar to the representation knowledge in artificial intelligence.

The exemplar theory is that concept is exemplar and the exemplar “is a body of knowledge about the properties believed to be possessed by a particular member of a class.” [5] (p. 93). Similar to the prototype, people can judge whether one thing belongs to one category by checking the similarity between the exemplar and the thing. The difference between exemplar and prototype is that the exemplar is an actual member of a category extracted from memory, but the prototype is an abstract average of category members [6]. People can make quick judgments using prototype, but when people encounter atypical category members, it’s much easier for them to use exemplar to make a judgment.

The theory is that “concepts stand in relation to one another in the same way as the terms of a scientific theory and that categorization is a process that strongly resembles scientific theorizing.” [7]. Unlike prototype and exemplar, the content of a concept in theory is determined by its use in the theory, rather than by its mere constituent parts. The theory represents the background of knowledge, which contain hidden essences, causal laws and functions [8]. Compared with prototype and exemplar, theory is more likely to represent the moral concept, since it contains more background knowledge.

In summary, prototype, exemplar and theory are three completely different heterogeneous concepts. The concept of the prototype stores statistical knowledge and involves cognitive processes based on linear similarity calculation. Exemplar concepts store knowledge about specific individual attributes and involve cognitive processes based on nonlinear similarity calculations. Theory stores knowledge of cause and effect, law and function, which is similar to optimal or causal reasoning [9]. Although the three concepts represent completely different cognitive processes, they all have one thing in common: they all have their own cognitive mechanism, which is based on representation. Prototype’s concept is much more similar to the concept learning that Brenden M. Lake is working on, and much closer to basic artificial intelligence that is currently being studied. The Theory concept is more closely related to the ethical concept known as safe autonomous AI studied by Kaj Sotala. Since artificial intelligence is built on a variety of cognitive mechanisms, and these three concepts exactly represent cognitive mechanisms, artificial intelligence has concepts in the paradigm of cognitive science.

2.3. From the Perspective of Philosophy

From some perspectives of philosophy, “concept” has transcendental color. This paper will illustrate what concept is based on two philosophers, who are Fodor and Hegel.

Fodor, for example, argues that “almost all of our words are innate. The main point of this argument is that lexical concepts are primitive-they lack structure-and the primitive concepts cannot be learned.” [10]. In Fodor’s opinion, having a concept of x is being able to have propositional attitudes about x as x. What are propositional attitudes? “People have beliefs, desires, opinions, wishes—what are called propositional attitudes in philosophy.” [5] (p. 32). Since the propositional attitude is only owned by humans, the concept is only owned by humans. From Fodor’s point of view, AI does not have concepts. Neither Brenden’s Bayesian program learning nor Kaj Sotala’s attempt to implant concepts into artificial intelligence can make artificial intelligence possess a propositional attitude and can make artificial intelligence possess beliefs and desires. Moreover, since our stock of words are innate, even the strong AI of the future cannot have concepts.

According to Hegel, concepts are divided into two levels. One level is ground-level concepts and practical concepts, which are mainly used to describe and explain empirical goings on. The other level is meta-concepts, which play a unique role in expressing the possibility of describing and explaining the characteristics of the framework. Meta-concepts are priori and independent of any specific use of ground-level concepts. Meta-concepts, as an expressive organ of self-consciousness, are the premise of self-consciousness [11]. Starting from Hegel’s thought, AI can only learn ground-level and practical concepts, but not meta-concepts. However, since meta-concepts are the basis of ground-level and practical concepts, simply learning the ground-level and practical concepts without learning meta-concepts is not learning real concepts. In this sense, AI does not have concept.

3. Conclusions

In summary, AI has concept in the sense of common sense and cognitive science but does not have concept in the sense of philosophy, which has transcendental color.