Conference material: "Proceedings of the International Conference on Computer Graphics and Vision “Graphicon” (19-21 September 2023, Moscow)"
Authors:Khaniev R.E., Diane S.A.K.
3D Approach to Visualization of Error Function in the Neural Network Tuning Problem
The report presents an approach to visualization and analysis of the effectiveness of a neural network using one-dimensional and three-dimensional graphs. A generalized block diagram of algorithmic support for solving the problems of configuring and visualizing neural network models is proposed. As part of the study of the influence of several macroparameters on the quality of network learning, particular one-dimensional dependencies of RMSE errors and a generalized three-dimensional error function are given, which takes into account the number of neurons in the hidden layer, the number of training examples and the number of training epochs. Software implementation of algorithms for calculation, training and visualization of neural networks is carried out in Python. The analysis of the obtained three-dimensional graph made it possible to conclude about the nonlinear nature of the obtained multidimensional dependence and to determine the best combination of macroparameters using the expert evaluation function.