The authors propose an approach that implements the Kalman filter's functionality based on generative neural networks and deep learning. Among the most plausible data processing techniques there are Kalman filters, characterizing time-varying phenomena. Kalman filters are used for intuitive probabilistic interpretation, they have a simple functional form and are widely used in various disciplines. Based on recent variational methods for studying deep generative models, the authors introduce a unified algorithm to efficiently study a wide range of Kalman filters. Of particular interest is the use of temporary generative models to derive dependencies without sufficient knowledge of the physics of the simulated phenomenon. The prerequisite for this work is the ever-deepening intellectualization of electronic devices, implying an expansion of the range of tasks solved by electronic devices and an increase in the number of electronic devices. The mentioned expansion imposes high requirements for creating complex and numerous software for electronic devices. The approach proposed by the authors demonstrates the possibility of reducing the requirements for creating new software. The authors show that a neural network can implement very different algorithms based on its functionality, and these algorithms are created not by a human programmer, but by the neural network itself. The result of the work of the Kalman neural network filter (created by the authors) is the direct creation of the final signal directly by the neural network, without human reprogramming.
Volume 12 | 03-Special Issue
Pages: 1513-1519
DOI: 10.5373/JARDCS/V12SP3/20201405