Information Integration Workshop ,95 (IIW-95)

Ryuichi Oka TRC Theory and Novel Function Research Department

In the RWC, we assume that information integration is a very useful approach to achieve "Flexible Information Processing." To discuss the many facets of this approach, we sponsored a 3-day "Information Integration Workshop IIW-95" at the Tsukuba Daiichi Hotel from April 26 through 28, 1995. Those papers with permission of the authors will be included in the RWC Technical Report (in English) at a later date. Please read the RWC Technical Report for further details.

The following describes the workshop's purpose and the discussion topics (provided by the program committee). I hope that vigorous discussions that suit the workshop's purpose described below will continue in the future.

Workshop Purpose

There is growing interest about "information integration" research, with a view to properly processing information in the real world. The intent of this workshop was to define information integration as the basic problem of intellectual information processing, to clarify the problem from the aspects of conceptualization, theory, models, algorithms and architecture, and to search for the direction of further development.

In the Real World Computing Program (RWCP), information integration is considered to be a powerful means to achieve "flexible information processing." We understand that the essence of human intellectual ability and new information processing principles will be clarified by the concept of integration. But discussions are needed as to how this is related to conventional artificial intelligence and pattern recognition.

Information integration means the integration of voice with image, of symbol or language information with pre-symbol or non-language information, and so on. By integrating information we would be free from limitations of search functions of a system, and expect an effect that cannot be realized with one type of information. For effective information integration we have to develop a good internal model.

In the workshop's preparatory stages and through ongoing discussions, our object has never been to impose on others our views on information integration, but rather to create a common understanding. To achieve this goal, we prepared a list of issues as topics of discussion. These issues are classified into the following five categories:

  1. Information representation for integration (internal model/internal representation and information integration)
  2. Integration of multimodal information (particularly, the relationship of symbols with information integration)
  3. Integration of micro and macro information (the relation of hierarchy/learning with information integration)
  4. Integration of agents and environment (environment, real time interaction and information integration)
  5. Software and architectures concerning integrated processing (computers and information integration)
  6. Other

List of issues

Issues concerning (A)

  1. Does it make sense to unify the representational forms of intemal models? If so, what form of representation would be good? An assembly of restrictive conditions? A statistical model? A chaos-dynamics system? Is there any evidence that a probabilistic model (statistical model) is suitable?
  2. Is real-time linked with a model suited to the application of a specific algorithm?
  3. In an internal model, does it make sense to distinguish "time" and "space"? (How about in an animal brain?)
  4. Is internal representation using temporal oscillation useful for information integration? (How about in an animal brain?)
  5. How is "integration" related to the representation operation for converting ill- posed problems in the real world into well- defined problems? And without such a conversion, insoluble at all? For example, can it be said that explicit models will disappear with an idea such as "integration of an environment with systems.
  6. How to deduce a higher model from a primitive model? Can local computation (proximity action) be used? What about foreseeing construction of a new mechanical Gestaltism for a "group of elements with structure"?

Issues concerning (B)

  1. At what level and density is multimodal information integrated? What sort of model gives fast and reasonable results? At what stage do media merge?
  2. What is a symbol in view of engineering usefulness? Is a symbol as internal representation essential for information processing?
  3. Should we build such an internal model to work by having existing symbols "correspond" to presymbols (patterns)? Or, should we build such an internal model to spontaneously produce symbols from presymbols?
  4. Related to the first question in (9), can an example be made where "corresponding" produces a substantial effect? First, what is a "happy situation" when symbols and patterns are integrated? Is there any qualitative merit acquired over and above a mutual complement or decreasing search range?
  5. By causing the "emergence" referred to in the second question in (9), can an example be made to produce a useful engineering effect? Is any universal theory possible for generating symbols?
  6. Is it possible to work out a good internal model to deal with symbols and algorithms? Is there a so-to-say semantic space to represent the mutual relations between symbols?
  7. If the problem of (12) cannot be solved directly, is it possible to extinguish the problem itself using any of the following positions? ・A position to negate an explicit internal mode ・A position to negate usefulness of symbol ・A position to spontaneously generate the par of an internal model which deals with symbols?
  8. (14) What is the relation of integration to the evaluation of "distance" and "place" as parameters which can distinguish a symbol and a pattern?
  9. For integration problems, is it useful to evaluate the relation of a symbol to a pattern using "compression of information" and "quantum"?
  10. Is it possible to distinguish the following?: ・Integration as a problem characteristic o specific media such as speech and image ・Integration as a common problem

Issues concerning (C)

  1. Is the hierarchy of a temporal scale, a space scale and an abstract level essential for intellectual information processing?
  2. Is the view correct that learning is a reintegration of strata?
  3. Are learning algorithms, i.e., "meta-algorithms" just a sort of methodology like "OR"?
  4. Is it possible to integrate the extent of the outer world by learning?
  5. What is the content in which self-organization is fundamentally different from clustering? How about its formulation?
  6. Is it proper to attribute the necessity of learning (or hierarchy) to improvement of forecasting ability and information compression ability?
  7. In actual applications, how significant is it to adjust parameters of an internal model automatically on behalf of man, To what degree is it feasible? Can substantial meaning be given to such metoric as "structural learning that is not an adjustment of parameters.
  8. Does it make sense to further increase the number of strata and consider such problems as learning of the way of learning or evolution of the way of evolving? Is it possible to make an engineering example to show the effectiveness of such a concept? (How about in living creatures?)

Issues concerning (D)

  1. How can there be an integration of the outer world and models in the "real world" by real- time interaction?
  2. What can be found in making mechanisms for robotic research?
  3. What are the features of information integration in a humanoid robot?
  4. How to harmonize and integrate a human motion with a robotic motion? In terms of artificial intelligence, what can be learned from a human motion?
  5. What is the core of the internal structure of an agent? Why?
  6. Just what part of the outside should an outer model in an agent grasp? What about the relation of symbols to patterns?

Issues concerning (E)

  1. Are there any unified internal models and algorithms that have value in making a single dedicated computer? Or, are individual special- purpose computers necessary, depending on the types of information or processing levels?
  2. From the standpoint of computer language, is it not possible to find a principle that links the worlds described in different systems? Is there no language system whose fundamental principle is linking?
  3. From the standpoint of computer language, is it not possible to reconsider the "real number," which is a representation of the geometric world (pattern)? How about related concepts, for example, "analysis" (in a sense of the theory of functions)? Is it not feasible to make a system, that will include a geometrically linked one (manifold), and yet surpass it?

    List of papers and authors presented at the "Information Integration Workshop" sponsored by RWC