Evaluation of Neural Network Output Results Reliability in Pattern Recognition

Abstract

Artificial neural networks (ANN) are well-known effective parallel systems which have successfully approved themselves in solving of complicated artificial intelligence problems. The practice of the widespread ANN application due to their high efficiency in solving non-formalized or hard-formalized problems associated with the need for ANN training on experimental material particularly in pattern recognition. In solving problems of pattern recognition the feedforward neural network is usually used due to the simplicity of algorithmic implementation, the availability of advanced training methods, the possibilities of multi-parallel computations. When neural network classifier implementing within decision support systems, it is necessary to assess the ANN results reliability based upon interpretation of the output signals to establish trust between users and the neural network algorithm. In this paper the ANN output results reliability evaluation method in terms of the degree of belonging of the recognized patterns to the originally specified classes is considered. The proposed method is based on the computation of Euclidean distance between the actual ANN output vector characterizing the class of the object recognition and a set of sample vectors defining a priori known classes at the training stage followed by its evaluation by the curve construction of the normal probability distribution law coinciding with the Gaussian function. A feature of this method is the construction of an individual probability distribution curve computed for each ANN output vector. An experimental research of the proposed method in MATLAB is presented on the example of solving the known Fisher’s Iris Database classification for input data without noise and with noise. The obtained results confirm the adequacy of the proposed method which can be used both in independent neural network pattern recognition systems and within decision support complexes.