Method of Logical Interpretation of Neural Network Solutions

This paper proposes a method for logical interpretation of neural network solutions. In order to logically interpret the operation of a neural network, various methods can be used that help visualize and analyze the internal processes occurring in the network. The approach under consideration examines only the input data and the results of the neural network solution and does not take into account the weights, structure, learning method. Using Boolean integro-differential calculus, it considers possible logical relationships between the input data and the results of the decisions. Combining these relationships, a function is obtained that allows you to analyze in detail the decision area of the neural network, find the most important features and hidden patterns in the data. This is especially useful for solving problems in which the exact form of the relationship between the input data and the result is unknown, but a sufficient amount of experimental data has been accumulated.