The workshop was held at the Tsukuba Daiichi Hotel on November 9 and 10, 1995. This report covers the objectives and Session (I) of this workshop.
[Objectives]
The RWC's Theory and Novel Functions research is characterized by "information integration". In the last workshop, discussions focused on the presentation of papers on more than 30 predetermined subjects.
This workshop was held to narrow the subjects and to clarify the core concepts of "information integration" based on the last workshop. The following three subjects were addressed:
(I) "Languages" for representing knowledge and information integration
(II) Non-linguistic communication -- how it should be understood and utilized
(III) Principle for constructing autonomous intelligent systems-environments, problems, and segmentation and integration of systems
These subjects were subdivided for specific discussions as shown below. For Subject (I), (a) how the representation capability of symbolic languages and the descriptive capability of programming languages should be used to enable computers and the real world to interact effectively, (b) what is required for (a), and (c) how far it has been achieved.
For Subject (II), (a) what is the current status of research and experience on multimodal dialogs, (b) what are the directions of methods for integrating patterns (non-linguistic) and symbols (linguistic), and (c) what form should the man-machine communication take, learned from man-man communication. For Subject (III), (a) what are the recent insights on humans' initial and intermediate visions (focused on segments and perception), (b) what does the autonomy (of robots) stand for, and (c) how is such autonomy acguired (architecture for autonomy).
The above subjects were discussed in invited speeches and panel discussions.
[Session (I)]
"Languages" for Representing Knowledge and Information
Integration
(Chairman: Hideyuki Nakashima, ETL)
In this session, information integration in terms of languages and symbol processing is discussed by both agreeing and disagreeing parties.
First, a speech entitled "Constraints and Integrated Computational Architecture", which was given by Koiti Hasida from ETL, was chosen as a basis for the discussion. The main points of his speech were that describing and executing all processings in natural language from the constraint point of view enables processing with a close interrelationship among all modules, which may thus allow the bottlenecks in natural language processing to be solved, and this has been implemented in programming languages.
Panel Discussion 1: Some thoughts on Languages
Some of the opinions expressed in this panel discussion are shown below:
Hideyuki Nakashima - I Consider the role of languages in information integration. Language is used to represent something. It is not the world (environment) itself, but serves humans. Because humans think in language, their thinking is affected by the structure and vocabulary of the language they are using. Integration only begins by capturing objects (world) with the help of things (symbols). That is why I maintain the importance of assessing how the world is understood by using programming languages.
Ryuichi Oka - Symbols are those that have already been perceived, and the descriptive languages apply local constraints to them. Or, humans can only create local constraints. In contrast, patterns are not perceived. Symbols and patterns maintain a one-to-one correspondence. Operations on symbols are always parallel to those on patterns with associated generation of patterns, and new symbols may be produced. Symbol processing is likely to incorporate this process.
Hideki Asoh - The roles of programming languages include facilitating the writing of programs and hiding things. If constraint-oriented assumptions are made to reduce what is written, programming languages require that both patterns and symbols should be handled, that a learning function should be included in the constraints, and that the hierarchy should be resolved.
Yukito Iba - I think the evolution of AI is a content issue rather than a programming language issue. When AI is divided into the mechanism and the computational theory, the latter should refer to the content. Why, then, should the mechanism be referring to the programming language, although it is useful to be able to hide the things?
Kiyoshi Akama - The functions of new languages are to call for ideas of, to automatically generate, to ensure the universality and correctness of, to build a new theory of, and to provide a field of experiment for algorithms. Computational theory in this context presents the relationship between the specification of problems and the algorithm. And the solution is attempted using declarative problem solving. I think pattern-symbol integration is solved in it.
Yutaka Ishikawa - Programming languages require universality and a wide market. Libraries are convenient for users but are not convenient for optimization and control abstraction. Abstraction and optimization should be pursued at the same time.
Nobuyuki Otsu - A stochastic structure is one way of incorporating constraints for integrated processing of patterns and symbols using the Bayes decision rule as a common basis. Learning should be the basis of integration.
[Discussion]
- Let's consider the common elements in pattern processing and symbol processing. They are found in meta-level descriptions . And the problem is how the constraints should be written. What is observed in humans? (Nakashima)
- It may be satisfied via constraint variables between the environment and computers. (Hasida)
- Patterns and symbols are integrated from the start. We need to clarify the structure. Symbol processing leads to pattern processing eventually. Symbols are discrete structures of signals. (Matsuyama, University of Kyoto)
- Pattern processing is location-dependent due to its spatial extent, symbol processing is not. Therefore, they can be distinguished by the presence or absence of handling of the location. (Oka)
- Does pattern processing involve anything other than matching ? (Yutaka Ishikawa)
- How about a continuous change in faces in CG ? (Nakashima)
- Can pattern processing exclude any processing not dependent on closeness even if it handles patterns? (Oka)
- Pattern processing is the processing of a large amount of numeric data. (Asoh)
- Pattern processing is the processing of horizontally smooth and continuous things. Some say it also includes processing which expands to and analyzes high-dimensional vector space. (Iba)
- Pattern processing is processing which uses knowledge of continuity or similarity. (Asoh)
- The point is closeness as well as smoothness. (Iba)
- Then, is symbol processing which handles closeness and smoothness pattern processing? (Ishikawa)
- Symbol processing is rigid while pattern processing is smooth. Nothing has achieved both at the same time. (Hasida)
- Pattern processing is processing which can perform differential operations. Symbol processing is based on a simple principle that two symbols are identical or not and offers the merit of combination. Is there any processing which satisfies both? (Nakashima)
- What are the problems with complicated values and complicated combinations? (Asoh)
- Analogy, which deals with the closeness of what is written with symbols. (Nakashima)
- It seems to me that analogy problems can be solved clearly. (Asoh)
- Does the purpose determine the representation and topology (closeness)? (Otsu)
- Pattern processing deals with groups, while symbol processing deals with individuals. (Akito Sakurai)
- Isn't it necessary to distinguish between converting symbols into patterns and patterns with intrinsic extent? (Iba)
- Integration includes both vertical and horizontal ones. Is there horizontal integration in your natural language processing, Dr. Hasida? (Masumi Ishikawa)
- Yes, it lies between syntax analysis and semantic processing, and between speech and language processing. (Hasida)
- What would change in your model if it were made multimodal, Dr. Hasida? (Oka)
- No single way could do it. Multiple algorithms and combinations of them have to be created. (Hasida)
- In that case, isn't there a problem if we assume that they are converted into symbols at the input stage? For body-language, it has to deal with those states where patterns and symbols are not clearly separated, doesn't it? (speaker unknown)
- Symbols and patterns differ only in data structure, so can it be solved by using a declarative method by putting a model of the data structure into computational space? (Akama)
- Then what is new in the programming methodology of RWC? First there should be things to be solved, then new languages are born from there. (Suwa)
- In my opinion, RWC should go beyond the paradigm of writing in languages. And the constraints have to be written? RWC should eliminate the necessity of writing them. (Asoh)
(Wording: Ryuichi Oka)