D. I. Miloserdov, PhD Student, e-mail: dmmil94@yandex.ru, St. Petersburg Institute for Informatics and Automation of the Russian Academy of Sciences, St. Petersburg, 199178, Russian Federation

Architectural Features of Neural Network Forecasting Software Systems with Continuous Training

In recent years, a method of neural network event forecasting has been developed, based on the use of a pair of recurrent neural networks with controlled elements. This method allows you to make predictions without interrupting training. However, for its full use, a reasonable software implementation is necessary. This study considers the problem of searching for a software architecture that implements the method of neural network forecasting with continuous learning. Offers an improved prediction method that significantly reduces the required amount of memory. A procedure for accelerated calculation of the weights of neural network synapses has been developed. To assess the effectiveness of the proposed architectural solutions, a comparative analysis of various variants of software implementations was conducted. In systems developed with the proposed innovations, the requirements for memory and computing resources are much lower than in software implementations of the prototype method. For example, the amount of memory required has decreased by an average of 15 times, and system initialization has taken 16 times less time. At the same time, the strategy of maximum memory saving in such systems proved to be unproductive compared to the combined approach. Based on the obtained comparison results, recommendations are given for the use and choice of architectures depending on the specific tasks facing the end user, and the hardware and software environment in which the forecasting systems are supposed to operate.
Keywords: recurrent neural network, forecasting, software implementation, costs optimization, hardware resources

P. 641–647