Optimization of convolutional neural network structure with self-configuring evolutionary algorithm in one identification problem

Computer technology development has opened great opportunities for researchers in the field of dynamic systems optimization offering new algorithms and ways of their combination into complex and powerful intelligent information technologies. A great interest in the field of machine learning has been given to dynamic objects identification and pattern recognition in computer vision systems in particular recently. Computer vision problems arise in various fields: industry, security and surveillance systems, data acquisition and processing systems, computer-human interaction systems, etc. Neural networks are widely and successfully used for solving machine learning tasks. Neural networks are computer models based on network of nerve cells of a living organism. Using classical neural network causes major difficulties for solving computer vision tasks as they require significant computational and/or time resources for learning as well as they lose important information about topology of the original data. For this kind of tasks a special type of neural network called convolutional neural network was developed. Convolutional neural network (CNN) is the part of subfield of machine learning called deep learning. CNN is used as the main technology in this paper. It allows to build complex hierarchies of features and perform objects identification based on them. Using of the pooling layers provides invariance to size of the image and concept of parameters sharing can significantly reduce the number of parameters that have to be adjusted and therefore save computational costs and time. Standard training method for neural network (back propagation) has certain weaknesses that can be partially eliminated by using a evolutionary optimization algorithm. The quality of the solution depends on the neural network structure, which can also be adjusted using evolutionary algorithms. In this paper, the self-configuring genetic algorithm SelfCGA is used for CNN's structure and weighting coefficients adjustment. Proposed system is tested on the task of age identification based on the person's photo.