Sensitivity Analysis for Decision Boundaries

  • Authors:
  • A. P. Engelbrecht

  • Affiliations:
  • Department of Computer Science, University of Pretoria, Pretoria, South Africa, e-mail: engel@driesie.cs.up.ac.za

  • Venue:
  • Neural Processing Letters
  • Year:
  • 1999

Quantified Score

Hi-index 0.00

Visualization

Abstract

A novel approach is presented to visualize and analyze decision boundariesfor feedforward neural networks. First order sensitivity analysis of theneural network output function with respect to input perturbations is usedto visualize the position of decision boundaries over input space. Similarly,sensitivity analysis of each hidden unit activation function reveals whichboundary is implemented by which hidden unit. The paper shows how thesesensitivity analysis models can be used to better understand the data beingmodelled, and to visually identify irrelevant input and hidden units.