Can the G Factor Play a Role in Artificial General Intelligence Research?
Total Page:16
File Type:pdf, Size:1020Kb
Can the g Factor Play a Role in Artificial General Intelligence Research? Davide Serpico 1 and Marcello Frixione 2 Abstract. In recent years, a trend in AI research has ficial systems that behave intelligently in narrow started to pursue human-level, general artificial intel- domains, namely, “narrow AI”. These kinds of ligence (AGI). Although the AGI framework is char- artificial systems can carry out domain-specific acterised by different viewpoints on what intelli- intelligent behaviours in specific contexts and gence is and how to implement it in artificial sys- are, thus, unable to self-adapt to changes in the tems, it conceptualises intelligence as flexible, gen- context as general-intelligent systems can do [2- eral-purposed, and capable of self-adapting to differ- ent contexts and tasks. Two important questions re- 4]. main open: a) should AGI projects simulate the bio- The realisation of a human-level artificial logical, neural, and cognitive mechanisms realising intelligence has seemed unfeasible to many the human intelligent behaviour? and b) what is the scholars until recent years. However, in the last relationship, if any, between the concept of general two decades, the AI community has started to intelligence adopted by AGI and that adopted by pursue the goal of a “human-level” artificial psychometricians, i.e., the g factor? In this paper, we general intelligence (AGI). This is attested by address these questions and invite researchers in AI several conferences, publications, and projects to open a discussion on the theoretical conceptions on human-level intelligence and related topics 1 2 and practical purposes of the AGI approach. [4-5]. Although these projects point to many dif- ferent directions to be followed by AGI re- 1. INTRODUCTION: THE AGI search, they represent a new movement towards HYPOTHESIS the concrete realisation of the original dream of a “strong AI”. The dream of the first generation of AI research- Two important movements intertwined with ers was to build a computer system capable of AGI emphasise the importance of the simulation displaying a human-like intelligent behaviour in of the human mind. The first, known as Biologi- a wide range of domains. Since human intelli- cally Inspired Cognitive Architectures (BICA), gence is highly flexible with respect to different aims to integrate many research efforts involved tasks, goals, and contexts, making the dream in creating a computational equivalent of the come true would have required developing a human mind. The second, which has been ini- general-purposed thinking machine. tially proposed during the First Annual Confer- In spite of some initial success (e.g., Newell ence on Advances in Cognitive Systems (Palo and Simon’s General Problem Solver [1]), the Alto, 2012), aims to achieve the goals of the attempts of researchers did not result in a do- original AI and cognitive science, that is, ex- main-general AI. What they achieved was, ra- plaining the mind in computational terms and ther, the development of highly specialised arti- reproducing the entire range of human cognitive abilities in computational artefacts [3]. 1 School of Philosophy, Religion and History of Science, As we mentioned, the AGI community un- University of Leeds. Email: [email protected] derstands general intelligence as the ability, dis- 2 Dept. of Classics, Philosophy and History, University of played by humans, to solve a variety of cogni- Genoa. Email: [email protected] Proceedings of the Society for the Study of Artificial Intelligence and Simulation of Behaviour 2018 D. Serpico & M. Frixione, 2018 tive problems in different contexts. Thus, nearly 2. EMULATING OR SIMULATING all AGI researchers converge on treating intelli- GENERAL INTELLIGENCE? gence as a whole: indeed, intelligence appears as a total system of which one cannot conceive one Human intelligence is defined by psychometri- part without bringing in all of it [5-6]. Ben cians as a domain-general cognitive ability, Goertzel [4] delineates the core AGI hypothesis namely, the g factor (see Section 3 for further as follows: the creation and study of a synthetic details). Wang and Goertzel [5] have rapidly intelligence with sufficiently broad scope and dismissed any connection between AGI research strong generalisation capability is qualitatively and the psychometric concept of general intelli- different from the creation and study of a syn- gence. According to them, projects in AI are not thetic intelligence with significantly narrower interested in the psychological description of scopes and weaker generalisation capability. human intelligence, if not in a weak sense. What is general intelligence? How can it be However, this conclusion seems to be, at implemented in artificial systems? In order to best, premature. Indeed, some attempts in AGI address these questions, it is necessary to open a research have encompassed a notion of intelli- discussion on both theoretical and practical is- gence that should be evaluated through the sues in AGI research. lenses of empirical findings. Since general intel- In this paper, we aim to clarify what rela- ligence represents to many psychologists, neuro- tionship exists, if any, between the concept of scientists, and geneticists the most important and human general intelligence and the AGI hypoth- well-studied aspect of human psychology [7, 8], esis. General intelligence was first conceptual- we cannot see any strong argument against the ised in the early twentieth century within psy- possible role of the g factor in (at least some) chometric research. Remarkably, as we shall research in AGI. Let us see why. show, psychometrics is quite a different kind of An AGI project can aim to either emulate or psychological science than the one traditionally simulate human intelligence. In the case of emu- tied to AI, that is, cognitive science. We shall lation, an artificial system will display a human- argue that AGI researchers cannot safely rely on like intelligent behaviour regardless of details 3 the psychometric concept of general intelligence about its realisation or implementation. In the and should, rather, look at intelligence as emerg- case of simulation, instead, an artificial system ing from several distinct biological and cogni- will display general intelligence not only at the tive processes. behavioural level but also at the mechanistic and In Section 2, we analyse different view- processing level. In other words, human-level points on whether AGI research should emulate artificial intelligence is realised by underlying or simulate human intelligence. Since many AGI mechanisms which are analogue to those realis- projects are inspired by psychological, neurosci- ing human intelligence. The former case likely entific, and biological data about human intelli- represents the notion of AGI that Wang and gence, scholars in AI should care about the psy- Goertzel [5] have in mind. Our targets are, in- chometric theory of general intelligence, its stead, examples of AGI research characterised promises and perils. In Sections 3 and 4, we by the latter approach. summarise the fundamental aspects of such a Before analysing this approach in more de- theory by emphasising the widespread disa- tails, it is worth considering that whether AGI greement about the existence of general intelli- gence. In Section 5, we outline important impli- 3 Here, behavioural assessments, such as Turing’s test and cations for contemporary research on Artificial Nilsson’s employment test, can address whether we have General Intelligence. achieved a human-level artificial intelligence: in brief, systems with true human-level intelligence should be able to perform human-like tasks [9]. Proceedings of the Society for the Study of Artificial Intelligence and Simulation of Behaviour 2018 2 D. Serpico & M. Frixione, 2018 projects should emulate or simulate human intel- ing AI architectures by enlightening the func- ligence, and the relationship between BICA and tioning of central aspects of intelligence such as AGI, are controversial topics. Franklin and col- learning, attention, and memory. leagues [3, 10] agree that an AGI agent may be A dialogue between neuroscience and AI successfully developed by using an architecture research seems to be largely welcomed within that is not biologically inspired. However, they the AGI community. A survey conducted by argue, the goals of AGI and BICA are essential- Muller and Bostrom [13] highlights how, ac- ly equivalent. Indeed, AGI hopes to solve the cording to many researchers, a human-level AI problem already solved by biological cognition, is to be achieved by means of research ap- namely, to generate adaptive behaviour on the proaches tying AI to neuroscience—e.g., Inte- basis of sensory input. Since biological minds grated Cognitive Architectures, Computational represent the sole examples of the sort of robust, Neuroscience, and Whole Brain Emulation. Of flexible, systems-level control architectures course, the commitment to the simulation of needed to achieve human-level intelligence, psychological, cognitive, and biological aspects copying after these biological examples—as of human intelligence is exerted in many ways. BICA projects do—represents a valuable strate- Let us see some examples. gy. 4 Some projects belonging to BICA and Wang [11] disagrees with this point of AGI’s agendas (e.g., SyNAPSE, HTM, SAL, view. According to him, in a broad sense, all AI ACT-R, ICARUS, LIDA, the ANNs, the Human projects take the human mind as the source