Convergence properties of perceptron learning with noisy teacher

  • Authors:

  • Kazushi Ikeda;Hiroaki Hanzawa;Seiji Miyoshi

  • Affiliations:

  • Nara Institute of Science and Technology, Ikoma, Nara, Japan;Nara Institute of Science and Technology, Ikoma, Nara, Japan;Kansai University, Suita, Osaka, Japan

  • Venue:
  • IScIDE'12 Proceedings of the third Sino-foreign-interchange conference on Intelligent Science and Intelligent Data Engineering
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

This paper analyzed convergence properties of an online learning method when teacher's signal includes noise in the thermodynamic limit. The learning curve was analytically derived using a statistical mechanical method and its validity was confirmed by computer simulations. In this case, the learning curve shows an overshoot phenomenon. In order to elucidate why and how it occurs in this case, the asymptotic analysis of dynamical systems was applied to the differential equations that expresses the dynamics of the learning curve and showed that the phenomenon results from the properties of the system matrix of the equations.