Recurrent neural networks with controlled elements in restoring frame flows
Keywords:
distorted frame flow, recovery, recurrent neural network, logical structure, control, evaluationAbstract
Introduction: Various interfering influences raise pressing problems of promptly restoring the flow of distorted frames,
remembering about the background and dynamics of the event measurement laws. The traditional methods of recovering flows of
distorted frames do not fully take into account the peculiarities of this process. Purpose: Exploring the possibilities of recurrent neural
networks with controlled elements for restoring frame flows. Results: It is proposed to evaluate the potential of a recurrent neural
network with controlled elements by the number of successful options for restoring a distorted sequence of frames. Evaluation of the
capabilities of such neural networks according to the introduced indicator showed their strong dependence on the type of network
structure and settings. Recurrent neural networks with spiral structures of layers work better. As the number of the turns in the helix
grows, the network capabilities also grow. Enhancing the capacity of a network to restore distorted frame flows is feasible if we replace
unipolar functions of the synapse weights by bipolar ones. A significant increase in the capabilities of the neural networks under study
is possible by controlling the neuron excitation thresholds in order to provide sequential rather than parallel elimination of various
errors. In contrast to the conventional neural networks, recurrent neural networks with controlled elements can adapt to changes in
№ 5, 2019 ИНФОРМАЦИОННО
УПРАВЛЯЮЩИЕ СИСТЕМЫ 17
ОБРАБОТКА ИНФОРМАЦИИ И УПРАВЛЕНИЕ
the laws inherent in frame flows, and implement controlled associative signal processing. Experiments have shown that these neural
networks can use associative connections for taking into account deep current experience in signal processing, and be successfully used
for restoring distorted frame flows.