This article is cited in 5 papers

Control in Technical Systems

Entropy dimension reduction method for randomized machine learning problems

Yu. S. Popkov^abc, Yu. A. Dubnov^acd, A. Yu. Popkov<sup>ae

a</sup> Institute for Systems Analysis, Russian Academy of Sciences, Federal Research Center “Informatics and Control”, Moscow, Russia
^b Braude College of Haifa University, Carmiel, Israel
^c National Research University “Higher School of Economics”, Moscow, Russia
^d Moscow Institute of Physics and Technology, Moscow, Russia
^e Peoples' Friendship University, Moscow, Russia

Abstract: The direct and inverse projections (DIP) method was proposed to reduce the feature space to the given dimensions oriented to the problems of randomized machine learning and based on the procedure of “direct” and “inverse” design. The “projector” matrices are determined by maximizing the relative entropy. It is suggested to estimate the information losses by the absolute error calculated with the use of the Kullback–Leibler function (SRC method). An example illustrating these methods was given.

Keywords: entropy, relative entropy, projection operators, matrix derivatives, gradient method, direct and inverse projections.

Presented by the member of Editorial Board: P. S. Shcherbakov

Received: 24.01.2018

DOI: 10.31857/S000523100002747-5

 English version: Automation and Remote Control, 2018, 79:11, 2038–2051

Bibliographic databases:

• 
• 
•