Here is a list of possible themes for
the Symposium on GC5 at AISB'06

CONTENTS: THEMES

• 1. Requirements
• 2. Empirical evidence and theories
• 3. Designs for Models/Explanations
  • 3.1. High-level (virtual machine) designs
  • 3.2. Implementation designs: neural/chemical/developmental
    • 3.2.a. Natural/biological mechanisms
    • 3.2.b. Artificial mechanisms
• 4. Philosophical/conceptual issues
• 5. Practical recommendations for managing GC5
• 6. Social/ethical implications

Theme descriptions:

• 1. Requirements

  What needs to be explained/modelled and how can we check that we have good requirements specifications?

  E.g. some people think that what needs to be explained about vision is simply how a depth-map is computed -- distance to contact in all directions. Others believe that the function of a visual system is to segment retinal images and recognise objects. Three more very different alternative requirements specifications are found in Marr, in J.J.Gibson and in the work of people who emphasise dynamical systems. It is arguable they have all done only a partial requirements analysis for vision (or more generally perception).

  Similar things can be said about what needs to be explained/modelled regarding: learning, motivation, affect in general, linguistic ability, reasoning, action control, consciousness, etc.

• 2. Empirical evidence and theories

  Work in other empirical disciplines, including biology (e.g. animal behaviour, evolutionary theories), neuroscience, linguistics, cognitive/clinical/developmental psychology, social sciences, that helps either to refine the requirements or support/contradict proposed models and explanations (theme 3).

  In particular work on empirical evidence from very young children, unusual humans (e.g. people with brain damage) or other animals can be useful in helping to avoid narrow thinking based on what normal adult humans (in our culture) can do.

  What other animals do may provide evidence about evolutionary precursors and about unnoticed sub-systems in human capabilities.

• 3. Designs for Models/Explanations

  • 3.1. High-level (virtual machine) designs

    proposed as meeting some subset of requirements (whether for explanation or for applications). [The 'top-down' approach.]

  • 3.2. Implementation designs

    for neural/chemical/developmental mechanisms proposed as capable of supporting the high level designs. [The 'bottom-up' approach.]

    • 3.2.a. Natural/biological mechanisms

    • 3.2.b. Artificial mechanisms

      This could include reports on current work in progress, and reports on current systems with analyses of how they are inadequate (they are ALL inadequate, in relation to the long term goals of GC5), and how those inadequacies may be overcome or reduced.

    [I.e. *merely* reporting on what some system has achieved is not appropriate for a GC5 symposium.]

• 4. Philosophical/conceptual issues

  Discussion of what is meant by ascribing various kinds of capabilities and processes to animals or machines, and whether conceptual categories currently in use (e.g. cognition, learning, intentionality, sub-symbolic, emotion, information-processing, consciousness) are confused and in need of refinement. Eg. do our current concepts allow us to discuss which organisms do or do not process information or what would be a good taxonomy of varieties of types of information-processing capability.

  [This is not the place for discussion of lucas/searle/penrose arguments as to whether AI is impossible. We are concerned with how it is possible, if it is. ]

• 5. Practical recommendations for managing GC5

  There is an important need for a project like this to have some sort of roadmap, even if it is regularly revised, and criteria for assessing progress. It is also important to resist fragmentation of effort and destructive rivalries. A number of proposals regarding this were put forward as part of the earlier discussion of GC5, but there is scope for further improvements.
  • Discussion of targets, milestones, in the context of a scenario-based research methodology
  • Extract from the IJCAI'05 tutorial booklet.
    Proposes a scenario-driven research methodology based on a partially ordered network of carefully analysed long term, medium term, and near-term scenarios, from which both requirements and milestones are derived, using a mostly backward-chaining research methodology.
  • Towards a Requirements Grid: A Conceptual Framework and Draft Tool for Generating Requirements and Scenarios

• 6. Social/ethical implications

  Many people think social and ethical implications should always be discussed as part of such a project. Often such discussions tend to be either very shallow or wildly speculative or based on some strong ethical bias against advances in AI, or some combination of all of those. Care will be needed to ensure that we do not fall into such traps.

  There are potentially profound applications that could follow from significant progress with GC5, not just in the obvious areas of building new smart robots and other machines, but in connection with applications arising from a deeper understanding of how humans work, including new ways of doing education, counselling, therapy, diagnosis of brain disorders and other kinds of mental disorders, and perhaps new forms of treatment. I am especially interested in ways in which understanding better what enables a toddler to grow up to be mathematician could revolutionise mathematics education in primary schools. But I suspect that discussing all this is premature.

Last updated: 28 Feb 2006
Maintained by Aaron Sloman