Working Group D: Models and Inference

Summary

First, the title of the working group was discussed. Models are automata, grammars, etc. There is a trade-off between the level of primitives and the complexity of recognition. For the statistical approach the problem of extrancting the best discriminating feature is solved, but not for the syntactic approach. There are some NP-completeness results about grammatical inference problems. So a major questions is, whether one has to start with a human made grammar, guess it and test it. Another suggestion is, restrict to a finite set of rules and try to find the best grammar within this finite search space. On the other hand one has not to take NP-completeness results into account if the relevant strings do not tend to grow very large. Otherwise one has chosen the wrong primitives or wrong relations between them. But how can one describe the essential features of the letter E e.g., taking into account, that.

is recognized by humand. Perhaps the theory of catastrophes can explain, why syntactic pattern recognition procedures are so sensible againts small phenomena. There are several existing successful pattern recognition systems, in which high-level primitives are used, which involve a lot of knowledge.