![]()
1.Introduction
To unfold the secret laws and relations of those high faculties of thought by which all beyond the merely perceptive knowledge of the world and of ourselves is attained or matured, is an object which does not stand in need of commendation to a rational mind.
~ George Boole, An Investigation of the Laws of Thought, 1854
∞
1.1Can Machines Have Human-Level Intelligence?
In 1950, Turing’s paper on Computing Machinery and Intelligence challenged scientists to achieve human-level artificial intelligence, though the term artificial intelligence (AI) was not officially coined until 1955, in the Dartmouth summer research project proposal by McCarthy, Minsky, Rochester, and Shannon.
Turing suggested that scientists could say a computer thinks if it cannot be reliably distinguished from a human being in an “imitation game,” which is now called a Turing Test. He suggested programming a computer to learn like a human child, calling such a system a “child machine,” and noted that the learning process could change some of the child machine’s operating rules. Understanding natural language would be important for human-level AI, since it would be required to educate a child machine and would be needed to play the imitation game.
McCarthy et al. proposed research toward computer systems that could achieve every feature of learning and intelligence. They proposed to investigate how computers could understand language, develop abstract concepts, perform human-level problem solving, and be self-improving. They planned to study neural networks, computational complexity, randomness and creativity, invention and discovery.
McCarthy proposed that his research in the Dartmouth summer project would focus on intelligence and language. He noted that every formal language yet developed omitted important features of English, such as the ability for speakers to refer to themselves and make statements about progress in problem-solving. He proposed to create a computer language that would have properties similar to English. The artificial language would allow a computer to solve problems by making conjectures and referring to itself. Concise English sentences would have equivalent, concise sentences in the formal language. McCarthy’s envisioned artificial language would support statements about physical events and objects, and enable programming computers to learn how to perform tasks and play games.
Turing’s 1950 paper concluded by suggesting two alternatives for developing machine intelligence. One alternative was to program a computer to play chess; the other was to create a child machine and teach it to understand and speak English.
The first approach, playing chess, was successfully undertaken by AI researchers, culminating in the 1997 victory of Deep Blue over the world chess champion Gary Kasparov. We4 now know that this approach only scratches the surface of human-level intelligence. It is clear that understanding natural language is far more challenging: No computer yet understands natural language as well as an average five-year-old human child. No computer can yet replicate the ability to learn and understand language demonstrated by an average child.
Though Turing’s paper and the Dartmouth proposal both stated the long-term research goal to achieve human-level AI, for several decades there were few direct efforts toward achieving this goal. Rather, there was research on foundational problems in a variety of areas such as problem-solving, theorem-proving, game-playing, machine learning, language processing, etc. This was perhaps all that could be expected, given the emerging state of scientific knowledge about these topics, and about intelligence in general, during these decades.
There have been many approaches, at least indirectly, toward the long-term goal. One broad stream of research to understanding intelligence has focused on logical, truth-conditional, model theoretic approaches to representation and processing, via predicate calculus, conceptual graphs, description logics, modal logics, type-logical semantics, and other frameworks.
A second stream of research has taken a bottom-up approach, studying how aspects of intelligence (including consciousness and language understanding) may emerge from robotics, connectionist systems, etc., even without an initial, specific design for representations in such systems. A third, overlapping stream of research has focused on “artificial general intelligence,” machine learning approaches toward achieving fully general artificial intelligence.
Parallel to AI research, researchers in cognitive linguistics have developed multiple descriptions for the nature of semantics and concept representation, including image schemas, semantic frames, idealized cognitive models, conceptual metaphor theory, radial categories, mental spaces, and conceptual blends. These researchers have studied the need for embodiment to support natural language understanding and have developed construction grammars to flexibly represent how natural language forms are related to meanings.
To summarize the current state of research, it has been clear for many years that the challenges to achieving human-level artificial intelligence are very great, and it has become clear that they are somewhat commensurate with the challenge of achieving fully general machine understanding of natural language. Progress has been much slower than Turing expected in 1950. He predicted that in fifty years people would commonly talk about machines thinking, and that this would be an educated opinion.
While people do informally speak of machines thinking, it is widely understood that computers do not yet really think or learn with the generality and flexibility of humans. While an average person might confuse a computer with a human in a typewritten Turing Test lasting only five minutes, there is no doubt that within five to ten minutes of dialog using speech recognition and generation (successes of AI research), it would be clear that a computer does not have human-level intelligence.
Progress on AI has also been much slower than McCarthy expected. In 2006 he gave a lecture in which he said he had hoped in 1955 that human-level AI would be achieved before many members of his audience were born.
Indeed, while many scientists continue to believe human-level AI will be achieved, some scientists and philosophers have for many years argued that the challenge is too great, that human-level AI is impossible in principle, or for practical reasons. Some of these arguments relate directly to elements of the approach of this thesis. Both the general and specific objections and theoretical issues will be discussed in detail, in Chapter 4.
In sum, the question remains unanswered:
How could a system be designed to achieve human-level artificial intelligence?
The purpose of this thesis is to help answer this question, by describing a novel research approach to design of systems for human-level AI. This thesis will present hypotheses to address this question and present evidence and arguments to support the hypotheses.
1.2Thesis Approach
Since the challenges are great, and progress has been much slower than early researchers such as Turing and McCarthy expected, there are good reasons to reconsider the approaches that have been tried and to consider whether another, somewhat different approach may be more viable. In doing so, there are good reasons to reconsider Turing’s and McCarthy’s original suggestions.
To begin, this thesis will reconsider Turing’s suggestion of the imitation test for recognizing intelligence. While a Turing Test can facilitate recognizing human-level AI if it is created, it does not serve as a good definition of the goal we are trying to achieve, for three reasons. First, as a behaviorist test it does not ensure that the system being tested actually performs internal processing we would call intelligent. Second, the Turing Test is subjective: A behavior one observer calls intelligent may not be called intelligent by another observer, or even by the same observer at a different time. Third, it conflates human-level intelligence with human-identical intelligence. Rather than create human-identical AI, we may wish to create human-like, human-level AI. These issues are further discussed in §2.1.1 and §2.1.2.
This thesis will propose a different approach 5 that involves inspecting the internal design and operation of any proposed system to see if it can in principle support human-level intelligence. This approach defines human-level intelligence by identifying and describing certain capabilities not yet achieved by any AI system, in particular capabilities this thesis will call higher-level mentalities, which include natural language understanding, higher-level forms of learning and reasoning, imagination, and consciousness.
Second, this thesis will reconsider Turing’s suggestion of the child machine approach. Minsky (2006) gave a general discussion of this idea, also called the ‘baby machine’ approach. He said the idea has been unsuccessful because of problems related to knowledge representation: A baby machine needs to be able to develop new ways of representing knowledge, because it cannot learn what it cannot represent. This ability to develop new forms of representation needs to be very flexible and general.
It is not the case that people have been trying and failing to build baby machines for the past sixty years. Rather, as noted above, most AI research over the past sixty years has been on lower-level, foundational problems in a variety of areas such as problem-solving, theorem-proving, game-playing, machine learning, etc. Such research has made it clear that any attempts to build baby machines with the lower-level techniques would fail, because of the representational problems Minsky identified.
What we may draw from this is that the baby machine approach has not yet been adequately explored, and that more attention needs to be given to the architecture and design of a child or baby machine, and in particular to the representation of thought and knowledge. This provides motivation for Hypothesis I of this thesis (stated in §1.4 below), which describes a form of the baby machine approach. This thesis will discuss an architecture for systems to support this hypothesis and will make some limited progress in investigation of the baby machine approach. Chapters 3 and 4 will analyze theoretical topics related to this architecture and discuss how the approach of this thesis addresses the representational issues Minsky identified for baby machines.
Next, this thesis will reconsider approaches toward understanding natural language, because both Turing and McCarthy indicated the importance of natural language in relation to intelligence, and because it is clear that this remains a major unsolved problem for human-level AI. Indeed, this problem is related to Minsky’s representational problems for baby machines, since the thoughts and knowledge that a human-level AI must be able to represent, and that a baby machine must be able to learn, include thoughts and knowledge that can be expressed in natural language.
Although McCarthy proposed in 1955 to develop a formal language with properties similar to English, his subsequent work did not exactly take this direction, though it appears in some respects he continued to pursue it as a goal. He designed a very flexible programming language, Lisp, for AI research, yet beginning in 1958 his papers concentrated on use of predicate calculus for representation and inference in AI systems, while discussing philosophical issues involving language and intelligence. In an unpublished 1992 paper, he proposed a programming language, to be called Elephant 2000, that would implement speech acts represented as sentences of logic. McCarthy (2008) wrote that the language of thought for an AI system should be based on logic, and gave objections to using natural language as a language of thought.
McCarthy was far from alone in such efforts: Almost all AI research on natural language understanding has attempted to translate natural language into a formal language such as predicate calculus, frame-based languages, conceptual graphs, etc., and then to perform reasoning and other forms of cognitive processi...
