Transcription is the process to make a score from sound. It is necessary in musicology field and also one of the interesting targets as application of AI. In the transcription system, the techniques of both of signal processing and symbol processing are required. Most of previous works have been devoted to the transcription for monophony, because of the difficulty of frequency separation in signal processing. We have developed a sophisticated peak frequency extraction method called "Interpolation method in complex spectra", which solved the problem of frequency separation. But it also made us find that the most difficult problem in transcription lay in symbol processing. The system has to know what is music to be transcribed. This paper describes a new transcription system for polyphonic sound based on the musical constraints. The process to extract notes from sound and the process to make score from notes are discussed in this paper. In the former step, the system uses the tonal model and the beat tracer. In the latter step, the system uses the knowledge of musical structure. We made some experiments for piano, guitar and shamisen. Experimental results show that the system can transcribe music of the instruments whose tonal attack is clear and that it is effective for popular light music.

In this paper, I deal with learning and recognition in vision and utilize character recognition as an example. Although many researches are done on character recognition, they are mainly interested in applications and can not answer many questions in vision. Why can human beings recognize shifted, magnified, reduced, and/or distorted characters? Why do children frequently write a mirror image of character?.… Machine learning, especially similarity-based learning, requires many instances of a concept. However, human beings can recognize even rather distorted characters by learning from only a few instances. In this paper, I propose a new pattern recognition model. This model can recognize shifted, magnified, reduced, and/or distorted characters by learning from a single instance for each character. Moreover, it can answer some of questions in vision. An input pattern contains k_1*k_1 elements, and each element has any strength of stimulus. A structure of an input pattern is generated by using relations ; calculation of relations among elements, generation of segment nodes by using the relations, calculation of relations among the nodes, generation of more general nodes by using the relations, calculation of relations among the nodes, and so on. This structure is used to recognize the pattern and a part of the structure is memorized with weighted attributes by learning.

Qualitative and quantitative reasoning can be applied to the parameter tuning process for the queueing network (QN) type system for improving its bottlenecks. This process consists of 3 phases, i. e., the diagnosis on bottlenecks and their sources, the estimation of the effects of the bottleneck improvement, and the determination of new parameters. On the basis of "qualitative reasoning" and "quantitative reasoning", 2 expert systems have been developed, i. e., BDES (Bottleneck Diagnosis Expert System) and BIES (Bottleneck Improvement Expert System) which can be applied to the 1st and 2nd phases, respectively. This paper shows the applicability of qualitative and quantitative reasoning for bottleneck improvement in queueing networks.