VISUALIZATION OF PROCESS OF TRAINING OF THE NEURAL NETWORK NEOCOGNITRON FOR CHARACTER RECOGNITION

Article Name:		VISUALIZATION OF PROCESS OF TRAINING OF THE NEURAL NETWORK NEOCOGNITRON FOR CHARACTER RECOGNITION
Authors:		A.V. Kugaevskikh (Russian Federation)
		The paper is recommended by program committee of 24th International Conference on Computer Graphics and Vision GraphiCon’2014.
Address:		A.V. Kugaevskikh a-kugaevskikh@yandex.ru Tyumen State University, Tyumen, Russia
Abstract:		In training of any neural network there is a problem of an assessment of quality of training that is defined by training of separate layers or neurons. In case of large number neurons it is difficult to trace flows of neurons activation. This paper shows the visualization mechanism of neural network training, the way of neurons activation allowing from input neurons through activation of features and before activation of the output neuron showing a class tag is provided. Approbation of the mechanism was realized on the example of training of a neural network neocognitron, recognizing characters of Ancient Egyptian language. Any character can be recognized only on its features distinguishing this character from others. Tracing of an activation path of communications between layers allows to reveal missing communications, excess features, to trace a signal between layers and to modify training parameters for their gain. Average quality of recognition of the constructed neural network makes 82% in case of total number of neurons about 700 thousands on 6845 classes.
Language:		Russian

Scientific Visualization