Exploiting label information to improve auto-encoding based classifiers Articles uri icon

publication date

  • December 2019

volume

  • 370

International Standard Serial Number (ISSN)

  • 0925-2312

Electronic International Standard Serial Number (EISSN)

  • 1872-8286

abstract

  • In this paper, we propose to provide more information to Stacked Denoising Auto-Encoding classifiers in order to increase their performance. Specifically, we use the output of an auxiliary classifier to extend the input to those machines, and carry out the layer-by-layer auto-encoding training considering the input recovering and the label errors by means of a convex combination whose parameter is selected by conventional cross-validation. Extensive experiments support the effectiveness of this proposal, showing that the resulting machines offer better results than standard designs in all the cases, as well as a reduced sensitivity to the design parameters. The main conclusion of this study plus a number of avenues for further research close this contribution.

keywords

  • auto-encoding; classification; convex combination; depth; label error