· In collaboration with Prof. O’Sullivan (Washington Univ. in St. Louis) we are developing a new feature extraction method for a recognition problem. This method quantifies the information content of finite data available for designing a recognition system and is based on an information theoretic criterion called Minimum Description Length (MDL). According to this method, the incompressible part of the data is announced to be a noise and is discarded and the other part of the data is encoded together with the probabilistic model assigned to it. The MDL method is well accepted in the information-theoretic community. This is one of the most powerful modern model estimation criteria. Our work is intended to bring the ideas of the MDL criterion to solve the problem of designing an optimal recognition system based on finite data.