1 International Research and Training Center for Information Technologies and Systems
of the NAS of Ukraine and MES of Ukraine, Kyiv, Ukraine
|
2 International Research and Training Center for Information Technologies and Systems
of the NAS of Ukraine and MES of Ukraine, Kyiv, Ukraine
|
Abstract. The article analyses risk-oriented formulation of pattern recognition and machine learning problems. Based on arguments from multicriteria optimization, a class of improper strategies is defined that are dominated by some other strategy. A general form of strategies that are not improper is derived. It is shown that some widely used approaches are improper in the defined sense, including the maximum likelihood estimation approach. This drawback is especially apparent when dealing with short learning samples of fixed length. A unified formulation of pattern recognition and machine learning problems is presented that embraces the whole range of sizes of the learning sample, including zero size. It is proven that solutions to problems in the presented formulation are not improper. The concept of minimax deviation recognition and learning is formulated, several examples of its implementation are presented and compared with the widely used methods based on the maximal likelihood estimation.
Keywords: pattern recognition, machine learning, short learning sample.