ITMO University Digest: # 3 Neural networks: interesting articles from ITMO University journals

Today in the digest ( first issue and second issue ) you will find a selection of scientific articles on neural networks published in different years in the journals of the ITMO University: starting with the properties and characteristics of neural networks of various types, possibilities for improving the quality and speeding up the operation of neural networks when solving other tasks, modeling of various processes of the human brain and ending with various practical applications of neural networks.

Characteristics and properties of neural networks

Formation of a training sample of a neural network . Researchers solve the problem of learning a neural network: it must be effective and should not take too much time. As a solution to the problem, the authors propose to use the method of mathematical planning of an experiment to obtain neural network training pairs and describe this process in detail.
')
Synthesis of optimal artificial neural networks . The authors are working on the question: how to create an algorithm that allows you to select such options for the structure and settings of the weights of the neural network to spend a minimum of time on network training and to obtain the optimal result. To solve the problem, scientists propose to use a modified genetic algorithm that allows you to find many optimal options for building a network.

The use of a radial basis neural network (RBFNN) for solving boundary value problems of mathematical physics . The author proposes to use parallelization in the technology of the software and hardware architecture of the learning algorithm RBFNN - the effectiveness of this approach is confirmed by an experiment comparing the duration of learning RBFNN on different CPUs and GPUs.

Dynamic characteristics of the neural network model of spatial memory . Researchers decide the question: how to build a model of the “memory” of artificial intelligence, which is close in quality to human. As a solution, it is proposed to use a heteroassociative neural network and the authors' developments in the field of assessing the correctness of the memorized material.

The implementation of the mechanisms of creative thinking of artificial intelligence . Scientists are exploring a number of issues related to the implementation of the mechanism for solving creative problems by a neural network with bidirectional connections formed by the Fourier holography method. Links to articles: 1 , 2 , 3 .

Neuro-fuzzy voltage regulator of the control object . Researchers are analyzing the possibility of combining fuzzy logic and neural networks into a neuro-fuzzy control system, which allows to increase the accuracy of the analysis and increase the speed of calculations, as well as create object management systems whose functioning algorithms are difficult to formalize using traditional mathematics.

Neural network mechanism for reducing cognitive dissonance . The authors are exploring the issue of the implementation of mechanisms that cause the appearance of phenomena similar to those that are inherent in the biological brain - in particular, the mechanisms for reducing cognitive dissonance.

Scopes of neural networks

Search for similar samples of malicious code (p. 301-305) . Researchers describe the implementation of an automated search engine for similar samples of malicious code based on a neural network.

Identification of the person . The authors consider the question: how can different types of neural networks be used in biometric identification systems, and choose the types of neural networks that are most effective for solving the problem.

Analysis of the flow of images in real time (pp. 348-353) . The material provides the basics of the concept of recognition of raster images using artificial neural networks - a description of the main approaches, as well as the methods used for the processes of analysis and forecasting when working with a neural network.

Determination of the flow rate of the bulk substance . The authors describe the mechanism for using a neural network to build a device for controlling the flow of bulk materials and the results of an experiment to create and use such a flow meter. Link to the article.

Creating an analog-to-digital converter (ADC) based on the neural network . Scientists propose the option of creating an ADC architecture based on a neural network, as well as a solution that makes it possible to achieve self-routing of ADC signals — all this makes ADC more reliable and more accurate.