Logic for the new AI

Minds and Machines

The often investigated future of human-machine relationships, ranging from cooperation partners at work, to social bots in care homes and personal assistants at home, is based on an often implicit technological requirement of robotics and artificial intelligence (AI): the ability of those machines to communicate with us in a form familiar and comfortable to us. Thus, those machines will have to learn how to communicate, either through intuitively understandable signs, text, or audio. This issue deals mostly with the latter. We may assume that machines 'speak', or rather, have to become speakers. However, this simple statement is laden with philosophical notions both about the conditions of what it means to 'speak', and thus become a speaker, and whether machines will ever be able to achieve these conditions. One may argue that the key challenge is to teach machines to use language according to its rules. However, what distinguishes 'natural' speakers from artificial ones? There is more to human language use than mere linguistic rule following: What do humans do additionally to follow the rules of language-and the requirements of communication in general-that machines are not currently and may not be able to? Could machines perform speech acts? If yes, which ones can be performed without the presence of underlying conditions such as human intentionality? Should machines count as agents, or should we reconstruct their actions as "quasiaction"? What could be functional equivalents to speech, and where exactly do they

Cognition, 1976

Abstractions and Embodiments: New Histories of Computing and Society, 2022

This chapter considers the history of automatic speech recognition research in the United States and its curious role in the imagining and implementation of computing as a communication medium. Defined by researchers in the 1950s as an effort to map the “meaning-measurement relation” between the abstractions of language and the material proper- ties of the voice, speech recognition was enlisted as a stage for competing theories in artificial intelligence and human-computer interaction. As such, it became not only a site where a particular understanding of speech and language was being formalized but one in which the boundaries of human and machine perception were being imagined, defined, and negotiated through communication technology.

With the increasing ubiquity of natural language processing (NLP) algorithms, interacting with “conversational artificial agents” such as speaking robots, chatbots, and personal assistants will be an everyday occurrence for most people. In a rather innocuous sense, we can perform a variety of speech acts with them, from asking a question to telling a joke, as they respond to our input just as any other agent would. However, in a stricter, philosophical sense, the question of what we are doing when we interact with these agents is less trivial, as the conversational instances are growing in complexity, interactivity, and anthropomorphic believability. For example, open domain chatbots will soon be capable of holding conversations on a virtually unlimited range of topics. Many philosophical questions are brought up in this development that this special issue aims to address. Are we engaging in a “discourse” when we “argue” with a chatbot? Do they perform speech acts akin to human agents, or do we require different explanations to understand this kind of interactivity? In what way do artificial agents “understand” language, partake in discourse and create text? Are our conversational assumptions andprojections transferable, and should we model artificial speakers along those conventions? Will their moral utterances ever have validity? This special issue of Minds and Machines invites papers discussing a range of topics on human-machine communication and the philosophical presumptions and consequences for developing, distributing, and interacting with speaking robots. We invite the submission of papers focusing on but not restricted to: - What are philosophically sound distinctions between speaking robots, unembodied chatbots, and other forms of artificial speakers? - What constitutes discourse participants, and can artificial speakers ever meet those requirements? - Can artificial speakers perform speech acts, and if yes, can they perform all speech acts humans can perform? Or do robots perform unique speech acts? - What kind of artificial agent can be capable of what kind of language or discourse performance: chatbots, robots, virtual agents,…? - What is the role of anthropomorphism in modelling chatbots as possible discourse participants? - What is the role of technomorphism in modelling human interlocutors as technical discourse participants? - What are the normative consequences of moral statements made by artificial discourse participants? - How will communicative habits between humans change by the presence of artificial speakers? - How can semantic theories explain the meaning-creation of artificial speakers? - Are normative conventions in human-human communication (politeness, compliments) relevant and transformable/transferrable to human-machine communication? - Are there – analogous to human-human communication – any communicative presuppositions in human-machine communication? To submit a paper for this special issue, authors should go to the journal’s Editorial Manager https://www.editorialmanager.com/mind/default.aspx Deadline to submit full paper: October 1st, 2020 First round of reviews: October 2nd – December 1st, 2020 Deadline to resubmit paper: December 15th, 2020 Second round of reviews: December 15th – December 31st, 2020 Deadline for final paper: December 31st, 2020 Publication of special edition: March 2021

Behind the somewhat pompous name of this paper hides a much more modest aim. It is an attempt at clarifying one of the major problem areas found in the field of Computer Assisted Language Learning (CALL), specifically the use of Automatic Speech Recognition (ASR). The aim of CALL is to produce tools aiding the acquisition of foreign languages and, long-term, to produce full-fledged, virtual, multi-modal language tutors. In essence, the teacher in the machine. Focusing specifically on Computer Assisted Pronunciation Training (CAPT) I intend to present a, hopefully, pedagogic picture of how the mind works in relation to adopting the sounds of a foreign language (L2), both perception- and productionwise, and illustrate the shortcomings of ASR in comparison. I also intend to present you with some rather unsubstantiated ideas on language learning which can hopefully generate a wholesome debate.

Proceedings of the 1975 workshop on Theoretical issues in natural language processing - TINLAP '75, 1975

IEEE Potentials, 2000

Few human endeavors can be viewed both as extremely successful and unsuccessful at the same time. This is typically the case when goals have not been well defined or have been shifting in time. This has certainly been true of Artificial Intelligence (AI). The nature of intelligence has been the object of much thought and speculation throughout the history of philosophy. It is in the nature of philosophy that real headway is sometimes made only when appropriate tools become available. For instance, the nature and behavior of physical objects was a major topic of philosophy. That is until the experimental method and the advent of calculus allowed for the development of Physics. Similarly the computer, coupled with the ability to program (at least in principle) any function, appeared to be the tool that could tackle the notion of intelligence. To suit the tool, the problem of the "nature" of intelligence was soon sidestepped in favor of this notion: If a probing conversation with a computer could not be distinguished from a conversation with a human, then "artificial" intelligence had been achieved. This notion became known as the "Turing test", after the mathematician Alan Turing who proposed it in 1950. This challenge quickly attracted the best computer scientists in a worldwide search for techniques and principles of what soon became known as the field of Artificial Intelligence. The early efforts focused on creating "general problem solvers" like, for instance, the Soar system (Newell, Laird and Rosenbloom) which attempted to solve problems by breaking them down into sub-goals. Conceptually rich and interesting, these early efforts gave rise to a large portion of the field's framework. Key to artificial intelligence, rather than the "number crunching" typical of computers until then, was viewed as the ability to manipulate symbols and make logical inferences. To facilitate these tasks, "AI languages" such as LISP and Prolog were invented and used widely in the field. That this quest never strayed far from rigorous mathematical underpinnings was both its strength and its limitation. Its strength was to open a new fertile area of computer science. Its limitation was that "real world" problems tended to be too complex for the limitations imposed by mathematical rigor and the constraints of logic and symbol manipulation. Therefore, much effort continued to be focused on "toy problems." One idea that emerged and enabled some success with real world problems was the notion that "most" intelligence really resided in knowledge. A phrase attributed to Feigenbaum, one of the pioneers, was "knowledge is the power." With this premise, the problem is shifted from "how do we solve problems" to "how do we represent knowledge." A good knowledge representation scheme could allow one to draw conclusions from given premises. Such schemes took forms such as rules, frames and scripts. It allowed the building of what became known as "expert systems" or "knowledge based systems" (KBS). These types of systems could indeed help in real world problems (the author led a project for the first expert system to aid astronauts in performing some scientific experiments. It was called PI-in-a-Box). The technology that ensued from expert systems gave rise to the first instance of an "Al industry." Consulting "Knowledge Engineers" and products (Shells) could take some of the drudgery out of building these types of systems. The enthusiasm of this time, however, masked an important shift that had been made by this technology: "Real world" solutions were obtained by keeping the system's focus extremely narrow and limited in scope. These systems were, and, to a large extent, remain extremely "fragile." That is, unexpected inputs or straying from the scope of the system could easily result in unexpected and erroneous results. The most difficult aspects of intelligence to incorporate appeared to be understanding a) one's limits of knowledge and b) the, unfortunately, elusive "common sense." The very usefulness and continuing success of these types of systems has also brought to light the fundamental limitation of the behaviorist model of intelligence. This model has difficulty coping with the fact that intelligence seems to reside in the ability to achieve one's expertise and to use it appropriately more than, or certainly in addition to, the expertise itself. Again, this realization shouldn't take away from the continuing improvements and successes in these types of systems. Model Based Reasoning has emerged as a powerful approach to diagnosis, and planning and scheduling systems have had much success as well. The point is that AI, now increasingly called "Symbolic AI," has produced a new branch of computer science. Along with it, powerful tools have been created for knowledge representation, symbol manipulation, searching and optimization. AI is alive and well. However, many opine that its picture of intelligence is too fragmented to represent a satisfactory model of cognition.

2008

let.kub.nl/people/bunt/index.stm www.loria.fr/~romary

Lecture Notes in Computer Science, 1998

In this paper I take the risk of predicting how the conflict between anti-symbolic Vs symbolic, emergentist Vs explicit approaches to cognition and action will be solved in the next decades. I do not believe in a 'paradigmatic revolution' and I argue in favour of a 'synthesis'. I will illustrate how a synthetic paradigm can be built through the notion of different levels of reality description and of scientific theory, and through their interconnections thanks to bridge-theories, cross-layered theories, and layered ontologies. I will provide several examples of bridge-theories and layered ontologies with special attention to agents and multi-agent systems. In particular I will examine the theory of the mental counterparts of social objects illustrating the mental facet of norms and of commitment; the grounding of social power in the personal power; the cognitive bases of organisations. I will propose a layerd approach to the notions of agent, delegation, communication, conflict, as they apply to different levels of agenthood. I will sketch the problem of emergence amo ng intelligent agents by exploring the problem of unplanned cooperation and social functions. I will conclude with the importance of the new "social" computational paradigm in AI, and the emergent character of computation in Agent Based Computing. Will the "representational paradigm"-that characterised Artificial Intelligence (AI) and Cognitive Science (CS) from their very birth-be eliminated in the 21th century? Will this paradigm be replaced by the new one based on dynamic systems, connectionism, situatedness, embodiedness, etc.? Will this be the end of the AI ambitious project? I do not think so. Challenges and attacks to AI and CS have been hard and radical in the last 15 years, however I believe that the next century will start with a renewed rush of AI and we will not assist to a paradigmatic revolutio, with connectionism replacing cognitivism and symbolic models; emergentist, dynamic and evolutionary models eliminating reasoning on explicit representations and planning; neuroscience (plus phenomenology) eliminating cognitive processing; situatedness, reactivity, cultural constructivism eliminating general concepts, context independent abstractions, ideal-typical models. I claim that the major scientific challenge of the first part of the century will precisely be the construction of a new "synthetic" paradigm: a paradigm that puts together, in a principled and noneclectic way, cognition and emergence, information processing and self-organisation, reactivity and intentionality, situatedness and planning, etc. [Cas98a] [Tag96]. AI is going out of a crisis: crisis of grants, of prestige, and of identity. This crisis was not only due-on my view-to exaggerated expectations and overselling of specific technologies (like expert systems) tout court identified with AI. It was due to the restriction of cultural interests and influence of the discipline, and of its ambitions; to the dominance either of the logicist approach (identifying logics and theory, logics and foundations) or of a mere technological/applicative view of AI (see the debate about the 'pure reason' [McD87] and 'rigor mortis'). New domains were growing as external and antagonistic to AI: neural nets, reactive systems, evolutionary computing, CSCW, cognitive modelling, etc. Hard attacks were made to the "classical" AI approach: situatedness [Suc87], anti-symbolism, reactivity [Bro89] [Agr89], dynamic systems, bounded and limited resources, uncertainty, and so on (on the challenges to AI and CS see also [Tha96]). However, by relaxing previous frameworks; by some contagion and hybridisation, by incorporating some of those criticisms; by re-absorbing as its own descendants neural nets, reactive systems, evolutionary computing, etc.; by developing important internal domains like machine learning and DAI-MAS; by important developments in logics and in languages; and finally with the new successful Agents framework, AI is now in a revival phase. It is trying to recover all the original challenges of the discipline, its strong scientific identity, its cultural role and influence. We may in fact say that there is already a neo-cognitivism and a new AI. In this new AI of the '90s systems and models are conceived for reasoning and acting in open unpredictable worlds, with limited and uncertain knowledge, in real time, with bounded (both cognitive and material) resources, interfering-either cooperatively or competitively-with other systems. The new password is interaction [Bob91]: interaction with an evolving environment; among several, distributed and heterogeneous artificial systems in a network; with human users; among humans through computers. The new AI and CS are-to me-only the beginning of a highly transformative and adaptive reaction to all those radical and fruitful challenges. They are paving the way for the needed synthesis and are starting the job. 1.1 The synthesis Synthetic theories should explain the dynamic and emergent aspects of cognition and symbolic computation; how cognitive processing and individual intelligence emerge from sub-symbolic or sub-cognitive distributed computation, and causally feedbacks into it; how collective phenomena emerge from individual action and intelligence and causally shape back the individual mind. We need a principled theory which is able to reconcile cognition with emergence and with reactivity: Reconciling "Reactivity" and "Cognition" We shouldn't consider reactivity as alternative to reasoning or to mental states [Cas95] [Tag96]. A reactive agent is not necessarily an agent without mental states and reasoning. Reactivity is not equal to reflexes. Also cognitive and planning agents are and must be reactive (like in several BDI models). They are reactive not only in the sense that they can have some hybrid and compound architecture that includes both deliberated actions and reflexes or other forms of low level reactions (for example, [Kur97]), but because there is some form of high level cognitive reactivity : the agent reacts by changing its mind: plans, goals, intentions. Also Suchman's provocative claims against planning are clearly too extreme and false. In general we have to bring all the anti-cognitivist claims, applied to sub-symbolic or insect-like systems, at the level of cognitive system 1. Reconciling "Emergence" and "Cognition" Emergence and cognition are not incompatible: they are not two alternative approaches to intelligence and cooperation, two competing paradigms. They must be reconciled:-first, considering cognition itself as a level of emergence: both as an emergence from subsymbolic to symbolic (symbol grounding, emergent symbolic computation), and as a transition from objective to subjective representation (awareness) (see later for example on dependence and on conflicts) and from implicit to explicit knowledge;-second, recognising the necessity for going beyond cognition, modelling emergent unaware, functional social phenomena (ex. unaware cooperation, non-orchestrated problem solving, and swarm intelligence) a lso among cognitive and planning agents. In fact, for a theory of cooperation and society among intelligent agents mind is not enough [Con96]. We have to explain how collective phenomena emerge from individual action and intelligence, and how a collaborative plan can be only partially represented in the minds of the participants, and some part represented in no mind at all [Hay67]. Emergent intelligence and cooperation do not pertain only to reactive agents. Mind cannot understand, predict, and dominate all the global and compound effects of actions at the collective level. Some of these effects are self-reinforcing and self-organising. There are forms of cooperation which are not based on knowledge, mutual beliefs, reasoning and constructed social structure and agreements. But what kind/notion of emergence do we need to model these forms of social behaviour? The notion of emergence simply relative to an observer (which sees something interesting or some beautiful effect looking at the screen of a computer running some simulation) or a merely accidental 1 Cognitive agents are agents whose actions are internally regulated by goals (goal-directed) and whose goals, decisions, and plans are based on beliefs. Both goals and beliefs are cognitive representations that can be internally generated, manipulated, and subject to inferences and reasoning. Since a cognitive agent may have more than one goal active in the same situation, it must have some form of choice/decision, based on some "reason" i.e. on some belief and evaluation. Notice that I use "goal" as the general family term for all motivational representations: from desires to intentions, from objectives to motives, from needs to ambitions, etc. By "sub-cognitive" agents I mean agents whose behaviour is not regulated by an internal explicit representation of its purposes and by explicit beliefs. Sub-cognitive agents are for example simple neural-net agents, or mere reactive agents.