Extraction of object hierarchy data from trained deep-learning neural networks via analysis of the confusion matrix
Roman Olegovich Malashin · Journal of Optical Technology · 2016
We studied the possibility of extracting object hierarchy information from a trained neural network by analyzing the errors obtained on a test sample using an approach based on singular value decomposition of the confusion matrix. Experiments indicate that the methods investigated in this paper can be used to obtain a tentative clustering of classes. In addition, we show that the number of connections within a fully connected layer of a convolutional neural network can be reduced without adversely affecting recognition accuracy for a test sample using locally connected layers. At the same time, however, our experiments did not show that a layer organization consistent with the object hierarchy led to any improvement of the results.