Повышение эффективности аппроксимационного метода решения обратных задач геоэлектрики на основе использования глубоких сетей сверточного типатезисы доклада
Аннотация:The paper considers the issues of increasing the efficiency of using theapproximation neural network (ANN) method for solving inverse problems (includingmultiobjectives), which are reduced to a nonlinear operator equation of the first kind (respectively,to a system of operator equations). The ANN method consists in constructing an approximateinverse operator of the problem using neural network approximation structures (MLP networks)based on a pre-built set of reference solutions for direct and inverse problems. An increase in theefficiency of constructing such structures in the work is achieved with the help of additionaltransformations that make up the layers of a convolutional neural network and include the use ofconvolutional filters, reducing the dimension of input data, physical and algorithmic complexing,etc. The compressed feature maps resulting from these transformations are adapted to the solvedinverse problem of high dimension, allow to reduce the approximation error of the inverse operatorand, as a result, to reduce the final error in the solution of the inverse problem obtained by the ANNmethod. The results of applying the improved design of the neural network are demonstrated on 2Dand 3D model and natural examples.