Constructing a method for the conversion of numerical data in order to train the deep neural networks

This paper analyzes known types of deep neural networks, the methods of their supervised training, training the networks to suppress noise, as well as methods for encoding data using images. It has been shown that deep neural networks are suitable in order to effectively solve classification problems, in particular for medical and technical diagnosing. Among the deep networks, the convolutional neural networks are promising because of their simple structure and application of common weights, which makes it possible for a network to separate similar features in different parts of images. Training a convolutional network may prove insufficient for some diagnosing tasks, which is why it is advisable to consider modifications to the training method using data encoding and training to suppress noise in order to obtain a better result. We have proposed a method for training a convolutional neural network using numerical data converted to bitmap images, which would improve the accuracy of a network when solving the problems on classification and which would make it possible to apply the convolutional neural networks and their advantages in image processing by using tabular data as input. In addition, the proposed method requires no additional changes to the structure of the network. The method consists of four stages – the normalization using a method of min-max, conversion of data into two-dimensional images applying the float or thermometric encoding methods, the generation of additional images with the distortion of input data, and the preliminary training of a deep network. The constructed method was implemented in software and investigated when solving a number of practical tasks. Results of solving the practical tasks on technical and medical diagnosing have shown the effectiveness of the method at small numbers of the resulting classes and training instances. The method could prove useful when diagnosing a defect at the early stages of its manifestation when the volume of training data is limited