|
Article information
2000 , Volume 5, ¹ 3, p.83-109
Senashova M.Y.
The methods of error estimations of signals in neural networks
Neural networks with a layered structure formed by the layers of standard neurons are considered. Errors arising under the technical realization of networks, noise and damages are studied. Proceeding from the condition that the vector of the network output signals should be calculated with the given accuracy, allowable errors probable for the signals of each network element are determined. Two types of error assessment (guaranteed interval estimations and root-mean-square estimation of errors) are used. It is shown that the estimates of allowable errors can be obtained by a special process "back propagation of accuracy". This process consists in functioning of network with the same system of connections, but from outputs to inputs and with the replacement of elements on dual. This duality differs from the one used in the classical method of calculating estimation gradients (back propagation of errors).
[full text] Classificator Msc2000:- *68T05 Learning and adaptive systems
- 92B20 Neural networks, artificial life and related topics
Classificator Computer Science:- *I.2.6 Learning
Keywords: neural network, signal, error estimation, accuracy
Author(s):
Senashova M Yu
PhD.
Position: Senior Research Scientist
Office: Institut Computing Simulation of SB RAS
Address: 660036, Russia, Krasnoyarsk, Akademgorodok
E-mail: msen@icm.krasn.ru
Bibliography link:
Senashova M.Y. The methods of error estimations of signals in neural networks // Computational technologies. 2000. V. 5. ¹ 3. P. 83-109
|
|
|
|
