Neural Networks Interpretation Improvement

Abstract

The paper is devoted to studying the issues of interpretability of neural network models. Particular attention is paid to the training of heavy models with a large number of parameters. A generalized approach for pretraining deep models is proposed, which allows achieving better performance in final accuracy and interpreting the model output and can be used when training on small datasets. The effectiveness of the proposed approach is demonstrated on examples of training deep neural network models using the MNIST dataset. The obtained results can be used to train fully connected type of layers and other types of layers after applying of flatting operation.