Study of a machine learning model based on a convolution neural network by explainability methods

Explainability methods are used to investigate a machine learning model based on the convolution neural network architecture. Class activation maps are calculated by applying algorithms based on forward and backward propagation of image tensors through the network components. A redundancy analysis of class activation maps and a statistical analysis of network weights during image propagation are also performed. The purpose of the work is aimed at increasing the explainability of internal processes of convolutional neural network functioning on the basis of the ResNet50 model. As a result, regularities and consequences reflecting the mechanisms of convolutional neural network operation when solving the problem of image classification are presented.

Authors: I. A. Utkin, D. S. Nagorny

Direction: Informatics, Computer Technologies And Control

Keywords: explainability, convolution neural network, class activation maps, method of principal components, statistical analysis methods, distribution density

---

View full article