Presentation 1995/7/27
Convergence Analysis of Recurrent Neural Network with Self-feedback loops
Yoshihiro Tomikawa, Kenji Nakayama,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Recurrent Neural Networks (RNN) with self-feedback loops can find the solution of a combinatorial optimization problem with less trial than that with no loop. However, the theoretical analysis of this type of network have not been sufficient. In this paper, the convergence property of the RNNs with feedback loops is analyzed. For this purpose, the eigenvalue analysis and the stability analysis are used. The eigenvalues of the connection weight matrix yield the macroscopic information about network dynamics. The stability of the network state space gives the microscopic information about the convergence to the vertecies. From these analyses, it is conformed that the convergence property to the quasi-optimal solution is improved by increasing the coefficients for self-feedback loops, which correspond to the diagonal elements of the connection weight matrix. However, it is also conformed that since the quasi-optimal solution is essentially different from the least energy solution, RNNs do not always gurantee the optimal solution.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Recurrent neural network / Self-feedback loops / Combinatorial optimization problem / Convergence property / Eigenvalue analysis / Stability analysis
Paper #
Date of Issue

Conference Information
Committee NC
Conference Date 1995/7/27(1days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English)
Chair
Vice Chair
Secretary
Assistant

Paper Information
Registration To Neurocomputing (NC)
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Convergence Analysis of Recurrent Neural Network with Self-feedback loops
Sub Title (in English)
Keyword(1) Recurrent neural network
Keyword(2) Self-feedback loops
Keyword(3) Combinatorial optimization problem
Keyword(4) Convergence property
Keyword(5) Eigenvalue analysis
Keyword(6) Stability analysis
1st Author's Name Yoshihiro Tomikawa
1st Author's Affiliation Graduate School of Natural Sci. & Tech. Kanazawa Univ.()
2nd Author's Name Kenji Nakayama
2nd Author's Affiliation Faculty of Eng. Kanazawa Univ.
Date 1995/7/27
Paper #
Volume (vol) vol.95
Number (no) 189
Page pp.pp.-
#Pages 8
Date of Issue