UK COMPUTING RESEARCH GRAND CHALLENGES
	Machines for intelligent ghosts?	Controversies in Cognitive Systems Research

Controversies in Cognitive Systems Research

This 'controversies' website started as a portion of that one, but soon grew large enough to be separated off. It will probably continue to grow indefinitely, and suggestions and criticisms are welcome. The location will shortly move to the euCognition web site, where a Controversies in Cognitive Systems Research section has been added to the euCognition wiki. This will be jointly edited by David Vernon and Aaron Sloman, to whom suggestions and criticisms should be sent.

CONTENTS

• Introduction
• How 'cognition' should be defined.
• Whether all mental and neural mechanisms should be regarded as dynamical systems, and modelled as such.
• The role of neural mechanisms.
• Symbol-grounding vs symbol-tethering
• Whether only bottom-up research and emergent phenomena can explain
• The Tabula Rasa Theory of Infant Learning
• Sensorimotor ontologies: Somatic vs Exosomatic
• Approaches to vision: image-based and scene-based
• Statistical vs structural models
• Concepts of causation required by intelligent systems: Humean or Kantian, or both?
• Do intelligent systems need emotions?
• What sort of architecture does an intelligent system need?
• What is Symbolic AI - should it be rejected? Is it needed?
• Is information always in the eye of the beholder?
• Should all learning and development be controlled by a single general purpose learning mechanism?
• Did evolution provide a collection of pre-packaged modules that determine major human characteristics?
• How did human language evolve?
• Do current claims about embodiment have any non-trivial content?

  ---

• REFERENCES
• APOLOGY

Introduction
Many controversies are associated with GC-5 research
Although it is hoped that researchers can in principle agree on the kinds of functionality that will need to be explained, it is not expected that they will agree initially on theories and concepts to be used or on which sorts of mechanisms and architectures are likely to work. However if they can collaborate for a few years on the task of specifying what needs to be explained (because much of it can be based on observation of humans and other animals, in various physical and social environments) that may be able to lay the foundation for eventual agreement on tests against which theories and models can be evaluated.

What follows is a provisional and incomplete list of topics on which there is disagreement that might one day be resolved (or reduced) by starting from agreed requirements (at least in the form of well organised, properly documented, analyses of competences that need to be modelled, explained or replicated in robots), in no particular order:

• How 'cognition' should be defined
  There are many different views on this -- as illustrated by the alternative definitions reported on the euCognition wiki

  It is arguable that disputes about how to define a particular widely used word or phrase are pointless. The more important task is to get clear about the logical geography and logical topography surrounding and underlying the space of alternative definitions.

• Whether all mental and neural mechanisms should be regarded as dynamical systems. and modelled as such.
  E.g. see The Dynamical Hypothesis in Cognitive Science by Tim van Gelder, and the commentaries. (Behavioural and Brain Sciences Journal, 1998) and the papers referenced here.

• The role of neural mechanisms.
  Whether only neural net models will suffice to explain the full range of capabilities GC-5 and other projects investigating minds and brains are concerned with.

  One of the common criticisms of computational neural net models is that they grossly oversimplify the ways in which real neurones work, ignoring the diversity of types of neurones the variety of kinds of organisations of networks, and the role of chemical processes involving neurotransmitters and hormones, for instance the role of opioids in learning claimed by Biederman and colleagues in Perceptual Pleasure and the Brain, Irving Biederman, Edward A. Vessel in American Scientist, 94, 3, May-June 2006. (Also accessible as html here.)

• Symbol-grounding vs symbol-tethering Whether symbol-grounding theory should be accepted, or whether it is just a reincarnation of the old concept empiricist theory refuted over two hundred years ago. E.g. see the opposing views on the euCognition wiki. A presentation linking symbol-grounding to concept empiricism is here (PDF)

  For more on Concept Empiricism, see: a recent paper by a philosopher attacking concept empiricism (PDF) Concept Empiricism: A Methodological Criticism (to appear in Cognition) by Edouard Machery, University of Pittsburgh Department of History and Philosophy of Science

  For a recent defence of Concept Empiricism see The Return of Concept Empiricism (PDF) [Penultimate draft of chapter in H. Cohen and C. Leferbvre (Eds.) Categorization and Cognitive Science, Elsevier (forthcoming). by Jesse J. Prinz Department of Philosophy, University of North Carolina at Chapel Hill

• Whether only bottom-up research and emergent phenomena can explain
  Whether the only way to build biologically inspired systems is to assemble low level components that faithfully model the behaviour and organisation of components of animal brains and see whether the required high level competences emerge, or whether high level functionality in a virtual machine can directly model important aspects of how human or other animal minds work.

  A currently popular view is expressed thus by David Cliff

  'Common to all of these new approaches was the observation that many naturally-occurring systems, at one level of analysis, can be described as being built from components that are individually "simple" and that interact with each other in relatively "simple" ways; yet at another level of analysis these systems exhibit some "complex" overall behaviour that is not readily predictable from the individual components.'

  Biologically-Inspired Computing Approaches To Cognitive Systems: a partial tour of the literature Dave Cliff, Hewlett Packard Research Laboratory. HPL-2003-11 2003

• The Tabula Rasa Theory of Infant Learning
  Whether systems with human-like intelligence have to start off knowing nothing at all about their environment (the 'tabula rasa' theory) so that everything they ever know is a result of very general learning mechanisms (e.g. reinforcement learning) operating on the experiences of the individuals, or whether a considerable amount of prior knowledge of some sort (possibly meta-level knowledge about how to find out facts about our particular type of environment) has to be innate. Some of the issues are summarised here by Juyang Weng. For a discussion about what may not be learnable without some innate high-level (meta-level) knowledge see Learning to see a set of moving lines as a rotating cube.

• Sensorimotor ontologies: Somatic vs Exosomatic
  This is the question whether all knowledge acquired by an intelligent system relates to sensorimotor contingencies (which would imply that the only available ontology is a 'somatic' ontology) or whether it is possible to develop an exosomatic ontology referring to 'external', entities, states, processes, relations, etc. (This is very closely related to the symbol-grounding dispute.) The conflict is described and the sensorimotor view (which is now extremely popular) is criticised in Sensorimotor vs objective contingencies and in this presentation 'Ontology extension' in evolution and in development, in animals and machines.

• Approaches to vision: image-based and scene-based
  Whether 'image-based' vision, based on detectable patterns in image arrays and their relations to other sensory patterns or patterns in motor signals will suffice for intelligent robots, or whether robot vision needs to make explicit use of a 'scene-based' ontology including 3-D components, kinds of materials, 3-D structural relations, etc. A useful collection of presentations at a 2006 symposium on object recognition is here. A related question is whether focusing on recognition is the right way to make progress on vision, given that we, and presumably many other animals, can see complex objects that we do not recognise, because they are novel. That leads to the more general question: what are the functions of vision? Is it possible to reach agreement on answers to that as a way of making progress, independently of agreeing on what sorts of mechanisms can implement those functions. A 24 year old survey of functions of vision is here. An influential paper by Barrow and Tenenbaum is H. Barrow, and J.M. Tenenbaum. 1978. Recovering intrinsic scene characteristics from images. In Computer Vision Systems, eds A. Hanson and E. Riseman, Academic Press. A book on the image-based approach is David Forsyth Computer Vision, A Modern Approach (2003). The web site includes several useful slide presentations. The preface states:

  'We see computer vision--or just "vision"; apologies to those who study human or animal vision--as an enterprise that uses statistical methods to disentangle data using models constructed with the aid of geometry, physics and learning theory.'

• Statistical vs structural models
  Whether statistical approaches to language understanding, vision, learning, will suffice to produce all the required competences, or whether there is a requirement for human-like machines also to be able to understand structures and their relationships, e.g. grammatical, geometrical, chemical, biological structures.

• Concepts of causation required by intelligent systems: Humean or Kantian, or both?
  Whether a 'humean' concept of causation based on observed or inferred statistical correlations suffices to describe all learning and reasoning about causation or whether there is a need to make use of a 'kantian' notion of causation in learning about what the environment is like and how things work. For a brief introduction to the difference see Two views of child as scientist: Humean and Kantian. For a more comprehensive introduction to current AI work on causation see Judea Pearl's web page and especially his book.

• Do intelligent systems need emotions?
  Whether intelligent systems need emotions. In recent years, especially since the publication of Damasio's book Descartes' Error: Emotion, Reason, and the Human Brain and shortly after that Rosalind Picard's Affective Computing, many AI researchers have claimed that emotions are required for intelligence and proposed mechanisms purporting to show how they perform that function. Several such papers can be found in Who Needs Emotions?: The Brain Meets the Robot edited by Jean-Marc Fellous and Michael A. Arbib. However a strongly dissenting voice is found in Arbib's own paper at the end of the book and another one in this presentation Do machines, natural or artificial, really need emotions? arguing that the popular view is based on fallacious arguments, confused concepts and wishful thinking. On the other hand, it is indisputable, as David Hume pointed out over two centuries ago, that some sort of motivation is a requirement for intelligence to be applied to doing something rather than nothing, and for selecting among possibilities. That seems to be a tautology rather than an empirical discovery that might be supported by neuroscience. Contrast the claim (in 1981) that Robots will have emotions not because they are needed, but because some emotions are a side-effect of mechanisms that are needed and constraints on implementations of those mechanisms.

• What sort of architecture does an intelligent system need?
  What sort of architecture (or mixture of architectures) human-like robots will need. Many different architectures have been proposed in the last three decades, but it is still unclear whether any of them is close to what will meet the requirements. In particular it is arguable that in humans the information processing architecture is not something fixed, but grows itself between infancy and later life, continually adding new components and new links between old components, as opposed to merely adding new information within a fixed architecture, as most current theories suggest. For some recent overview papers on 'cognitive architectures' see :
  • Cognitive architectures: Research issues and challenges by Langley, P., Laird, J. E., & Rogers, S. (2006).
  • A Survey of Artificial Cognitive Systems: Implications for the Autonomous Development of Mental Capabilities in Computational Agents, D. Vernon, G. Metta, G. Sandini (2007), accepted for publication in the IEEE Transactions on Evolutionary Computation, Special Issue on Autonomous Mental Development.
  • A Framework for Comparing Agent Architectures Aaron Sloman and Matthias Scheutz, Proceedings UKCI'02, UK Workshop on Computational Intelligence, 2002

• What is Symbolic AI - should it be rejected? Is it needed?
  (For an analysis of some of the requirements that are relevant see Requirements for a Fully Deliberative Architecture).

  Some aspects of this controversy are befuddled by the fact that many people start from historical claims that are at least debatable. E.g. there is an online introduction to neural nets which offers a very clear example of such a characterisation of Symbolic AI, listing arguments that are frequently presented, but which are mistaken.

  The quotation is from this section of Chapter 1 of an early online draft of Kevin Gurney's Introduction to Neural Networks Published by Routledge in 1997. The section is quoted with the author's kind permission. The final version of this discussion in Chapter 11 of the published book proposes that both symbolic and neural mechanisms are needed because the different mechanisms serve different functions.

  Responses to the statements in the quoted text have been inserted [in square brackets].

  (Symbolic AI assumes that) it is possible to process strings of symbols which obey the rules of some formal system and which are interpreted (by humans) as 'ideas' or 'concepts'.

  [Many AI programs, even in the 1960s, made use of information both in the form of input (e.g. image features) and in intermediate structures that had no correspondence with human ideas or concepts.]

  It was the hope of the AI programme that all knowledge could be formalised in such a way: that is, it could be reduced to the manipulation of symbols according to rules and this manipulation implemented on a von Neumann machine (conventional computer).

  [It was assumed that this was how researchers would implement their theories. When the programs ran they could model many different kinds of computation, including things that are very different from von Neumann machines, including theorem provers, pattern matchers, spreading activation networks, production system interpreters, feature detectors operating on images or other forms of input.

  Likewise the vast majority of neural net researchers write their programs to run on von Neumann machines but when they run the programs implement non-VN virtual machines.]

  We may draw up a list of the essential characteristics of such machines for comparison with those of networks.

  • The machine must be told in advance, and in great detail, the exact series of steps required to perform the algorithm. This series of steps is the computer program.

    [This completely ignores work on planners, theorem provers, and learning systems which work out for themselves what to do.]

  • The type of data it deals with has to be in a precise format - noisy data confuses the machine.

    [There was a lot of AI research even on noisy images and speech and various techniques were developed for dealing with noise, e.g. use of hough transforms.]

  • The hardware is easily degraded - destroy a few key memory locations and the machine will stop functioning or 'crash'.

    [A neural net program will also crash if you destroy memory locations where the compiled code resides or where the data-structures used by the compiled code are located. However that fragile implementation may support a very robust virtual machine. Similarly many AI systems which are not implemented as linear algorithms but as collections of interacting components, e.g. rule-based systems, can be very robust.
    (E.g. Nilsson's work on teleoreactive systems. These constantly sense the environment and modify their current plans and goals accordingly.)
    Of course, all existing systems have limitations and flaws of various kinds. There is no evidence that any one approach solves all the problems. ]

  • There is a clear correspondence between the semantic objects being dealt with (numbers, words, database entries etc) and the machine hardware. Each object can be 'pointed to' in a block of computer memory.

    [This completely ignores the use of virtual machines containing items that do not have any specific location in physical memory. That important fact was also ignored by Newell and Simon, who generated much confusion by their 'Physical symbol system' hypothesis. They failed to make clear that they were talking about physically implemented virtual machine symbols, not about physical symbols.]

  There are some researchers who attempt to integrate both the concerns of theories about symbolic mental processes and sub-symbolic neural and other processes. Examples include
  • Arnold Trehub, 1991, The Cognitive Brain, MIT Press, Cambridge, MA,
  • Marvin Minsky The Society of Mind, William Heinemann Ltd., 1987, London
  • Marvin Minsky The Emotion Machine: Commonsense Thinking, Artificial Intelligence, and the Future of the Human Mind, Barnes and Noble, 2006.
  • Ron Sun, Duality of the Mind. Lawrence Erlbaum Associates., Mahwah, NJ:. 2002. (Also see the CLARION web site.)

• Is information always in the eye of the beholder?
  (Draft entry: Placeholder)
  One of the old questions is what it means to say that something has thoughts, beliefs, percepts, intentions, concerns, puzzles, memories, etc. -- all states with intensional content, or intentionality. Answers that have been proposed include: that it's an unanalysable aspect of being conscious; that it can be a result of fairly straightforward operations of symbol manipulating mechanisms in a machine that acts in an environment; that it cannot occur except in biological mechanisms; that it requires an evolutionary history to give an animal or machine semantic content (Millikan); that it requires rationality (Newell, Simon, Dennett), that it is not question of fact whether something has those features, merely a question of the stance adopted by the beholder (Dennett's 'intentional stance').

  A different sort of answer proposes that we consider information as being a kind of basic stuff in the universe, like, matter (of various forms), energy (of various forms), none of which is explicitly definable in terms of simpler, or less obscure concepts, though each of them is implicitly defined by their roles in the best explanatory and predictive theories available (now and in the future). That answer is elaborated here.

  That seems to be the answer implicitly assumed by many psychologists, neuroscientist, biologists and engineers who talk about information including Eva Jablonka & Marion J. Lamb in Evolution in Four Dimensions: Genetic, Epigenetic, Behavioral, and Symbolic Variation in the History of Life MIT Press, 2006. Selected for BBS book commentary, with precis here.

• Should all learning and development be controlled by a single general purpose learning mechanism?
  (Draft entry: Placeholder)
  It is clear that biological evolution, insofar as that is a single process, is capable of producing hugely varied results, as it has done on planet earth. But time-scales required for that to work are enormous. That raises the question whether there is an equally general mechanism, or possibly something less general but still very powerful, that can operate within the lifetime of an individual so as to produce whatever form of learning is required by that individual. In particular, is there a single learning mechanism that explains all of human and animal learning? (Compare the 'Tabula Rasa' theory which assumes that learning starts from complete ignorance about the environment).

  Various researchers in cognitive science and AI have attempted to identify such general mechanisms, e.g.

  • Reinforcement learning,,
  • Jeff Hawkins' HTM mechanism
  • The Goedel Machine.
  (How many others have been proposed? ART??, SOAR??, Edelman's 'Neural Darwinism'??)

  An alternative view is that different animals, and different sorts of robots will need to use different forms of learning, and that humans have discovered which sorts work well in different contexts and pass them on through various kinds of educational procedures. Some have claimed that there are different learning styles used by different human learners (which should also be recognised by teachers).

• Did evolution provide a collection of pre-packaged modules that determine major human characteristics?
  (Draft entry: Placeholder)
  See
  • Leda Cosmides and John Tooby Evolutionary Psychology: A Primer

    "To solve the adaptive problem of finding the right mate, our choices must be guided by qualitatively different standards than when choosing the right food, or the right habitat. Consequently, the brain must be composed of a large collection of circuits, with different circuits specialized for solving different problems. You can think of each of these specialized circuits as a mini-computer that is dedicated to solving one problem. Such dedicated mini-computers are sometimes called modules. There is, then, a sense in which you can view the brain as a collection of dedicated mini-computers -- a collection of modules. There must, of course, be circuits whose design is specialized for integrating the output of all these dedicated mini-computers to produce behavior. So, more precisely, one can view the brain as a collection of dedicated mini-computers whose operations are functionally integrated to produce behavior.
    ....
    Biological machines are calibrated to the environments in which they evolved, and they embody information about the stably recurring properties of these ancestral worlds."

  • Annette Karmiloff-Smith Beyond Modularity MIT Press, 1992, and . The BBS precis states:

    Beyond Modularity attempts a synthesis of Fodor's anti-constructivist nativism and Piaget's anti-nativist constructivism. Contra Fodor, I argue that: (1) the study of cognitive development is essential to cognitive science, (2) the module/central processing dichotomy is too rigid, and (3) the mind does not begin with prespecified modules, but that development involves a gradual process of modularization. Contra Piaget, I argue that: (1) development rarely involves stage-like domain-general change, and (2) domain-specific predispositions give development a small but significant kickstart by focusing the infant's attention on proprietary inputs. Development does not stop at efficient learning. A fundamental aspect of human development ("Representational Redescription") is the hypothesized process by which information that is IN a cognitive system becomes progressively explicit knowledge TO that system. Development thus involves two complementary processes of progressive modularization and rendering explicit.

  • Jackie Chappell and Aaron Sloman Natural and artificial meta-configured altricial information-processing systems, to appear in IJUC.
  • Jablonka and Lamb on evolution, referenced above

• How did human language evolve?
  • Did it start as a simplified communication system based on sounds or gestures, then become gradually more complex, while evolution of required brain mechanisms happened at the same time?
  • Was there first some rich, expressive, complex internal form of language required for kinds of perception, planning, predicting, manipulating, that occur in non-linguistic animals (e.g. primates, nest-building birds, some hunting mammals, whose resources were later recruited for external communication?
  • Is there a useful concept of language that does not involve both communication and mappings between semantic contents and external behaviours?

    SOME REFERENCES ON LANGUAGE EVOLUTION
    (This is just a tiny subset, not representative.)

    Marc D. Hauser, Noam Chomsky, W. Tecumseh Fitch (2002),
    The Faculty of Language: What Is It, Who Has It, and How Did It Evolve?
    Science 22 November 2002: Vol. 298. no. 5598, pp. 1569 - 1579

    Steven Pinker and Ray Jackendoff, (2005) The faculty of language: what's special about it?
    Cognition, 95(2):201--236.
    Also here

    (Reply to the above)
    Fitch, W. T., Hauser, M. D., and Chomsky, N. (2005)
    The evolution of the language faculty: Clarifications and implications. Also here.
    Cognition, Volume 97, Issue 2, September 2005, Pages 179-210

    Hajime Yamauchi (2004) Baldwinian Accounts of Language Evolution PhD thesis (PDF),
    Theoretical and Applied Linguistics, University of Edinburgh

    Michael Arbib M. (2005). From monkey-like action recognition to human language: An evolutionary framework for neurolinguistics.
    Behavioral and Brain Sciences, 28(2):105--124.

    Preprint available
    http://www.bbsonline.org/Preprints/Arbib-05012002/

    Beth Azar (2005)
    How mimicry begat culture
    APA Online Monitor on Psychology, Volume 36, No. 9 October 2005
    (Includes a few more references)

    Michael C. Corballis,
    From Hand to Mouth: The Origins of Language
    Princeton University Press 2003

    Review by James Hurford

    Aaron Sloman, (March 2007)
    What is human language? How might it have evolved?
    PDF Seminar presentation, arguing that if we define a notion of g-language (generalised language) to refer to forms of representation of information that include structural variability, some systematicity, compositional semantics (usually context sensitive), then internal g-languages evolved before human (external) language, are used by non-linguistic animals and pre-linguistic children, and explain both why sign languages are so easily learnt and how a community of deaf children could create an entirely new sign language. (Comments invited). Builds on The primacy of non-communicative language (1979)

• Do current claims about embodiment have any non-trivial content?

  Obviously being embodied is important when you are running, climbing trees, catching things, building a house, avoiding being eaten, working out whether you can reach something, sexually attracted, enjoying making music, and most of the time when growing from infancy to adulthood. But how important is it when you are lying flat on your back solving a mathematical problem, reading a book on ancient history, studying quantum mechanics, wondering how languge evolved? Is it perhaps more important that your ancestors were embodied?

  See the proceedings of the euCognition workshop http://www.eucognition.org/embodying_cognition_2006.htm
  EMBODYING COGNITION: TOWARDS AN INTEGRATIVE APPROACH?
  Palma de Mallorca, 14-16 December 2006
  especially the excellent paper on The Workshop Theme by Toni Gomila and Paco Calvo.

  There was also A special issue on situated and embodied cognition in Cognitive Systems Research Volume 3, Issue 3 , September 2002, Pages 271-274
  Edited by Tom Ziemke.
  The controversies about embodiment are closely related to several other controversies in this web site, including:

  • How 'cognition' should be defined.
  • Whether all mental and neural mechanisms should be regarded as dynamical systems, and modelled as such.
  • The role of neural mechanisms.
  • Symbol-grounding vs symbol-tethering
  • Whether only bottom-up research and emergent phenomena can explain
  • Sensorimotor ontologies: Somatic vs Exosomatic
  • Statistical vs structural models
  • Concepts of causation required by intelligent systems: Humean or Kantian, or both?
  • What sort of architecture does an intelligent system need?
  • What is Symbolic AI - should it be rejected? Is it needed?
  • Is information always in the eye of the beholder?

REFERENCES

• Wikipedia Entry on Cognitive Architecture
• Wikipedia Entry on Cognitive Science
• The AAAI AITopics web site
• Definitions of Cognition reported on the euCognition wiki
• Cognition Briefings (on euCognition wiki)
• euCognition Research Roadmap Project
• Cognitive architectures: Research issues and challenges by Langley, P., Laird, J. E., & Rogers, S. (2006).
• A Survey of Artificial Cognitive Systems: Implications for the Autonomous Development of Mental Capabilities in Computational Agents, D. Vernon, G. Metta, G. Sandini (2007), accepted for publication in the IEEE Transactions on Evolutionary Computation, Special Issue on Autonomous Mental Development.
• Aaron Sloman and Matthias Scheutz, A Framework for Comparing Agent Architectures, in Proceedings UKCI'02, UK Workshop on Computational Intelligence 2002.