Method for searching outlier objects using parameters of learning instability

The paper describes a new method of outliers detection in pattern recognition tasks. The authors define an outlier as an object which deviates significantly from the other objects of the same class. The method is based on simultaneous use of evaluated object estimates for classes and integral distortion of recognition algorithm that is caused by evaluated object. Usefulness of the developed technique was shown for the task of predicting if an inorganic compound of composition $A^{+3}B^{+3}C^{+2}O_4$ is formed under ordinary conditions. The method may be used for erroneous observations detection that is aimed to improve training information in different recognition tasks.