Analyzing decision boundaries of neural networks

Chulhee Lee, Eunsuk Jung

Research output: Contribution to journalConference article

Abstract

In this paper, we analyze decision boundaries of 3 layer feedforward neural networks that use the sigmoid function as an activation function. By analyzing the decision boundaries in the space defined by the outputs of the hidden neurons, we found that the decision boundaries are always linear boundaries and that the decision boundaries are not completely independent. We found that for a 3-pattern class problem, the decision boundaries in the space defined by the outputs of the hidden neurons should meet at the same intersection. And this dependency of decision boundaries is extended to multiclass problems, providing valuable insight into decision boundaries. In particular, for a K-pattern classes problems, we found that there are only K-1 degree of freedoms in drawing decision boundaries in the space defined by the outputs of the hidden neurons, though there are KC2 decision boundaries. Finally, we present some interesting examples of decision boundaries of neural networks.

Original languageEnglish
Pages (from-to)217-226
Number of pages10
JournalProceedings of SPIE - The International Society for Optical Engineering
Volume4113
DOIs
Publication statusPublished - 2000 Dec 1
EventAlgorithms and Systems for Optical Information Processing IV - San Diego, CA, USA
Duration: 2000 Aug 12000 Aug 2

All Science Journal Classification (ASJC) codes

  • Electronic, Optical and Magnetic Materials
  • Condensed Matter Physics
  • Computer Science Applications
  • Applied Mathematics
  • Electrical and Electronic Engineering

Fingerprint Dive into the research topics of 'Analyzing decision boundaries of neural networks'. Together they form a unique fingerprint.

  • Cite this