A Proposal of Neural Networks with Intermediate Outputs

Billy Peralta, Juan Reyes, Luis Caro, Christian Pieringer

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

The automatic data classification is an essential problem in machine learning, and it applies to different contexts such as people detection, health or astronomy. In recent years, deep neural networks have gained extensive attention due to their excellent performance on large and complex datasets. A neural network is a supervised method for classification, therefore typically requires a set of inputs and targets for the training process. However, it is possible to include auxiliary outputs that characterize aspects of the object of interest, which can accelerate the learning process. For example, in an image, a person may have extra outputs like attributes given by the presence of a hat or beard. However, the classical neural networks do not consider the presence of explicit auxiliary outputs. Furthermore, these outputs might be at a lower semantic level. We propose a framework that allows for using auxiliary outputs connected to hidden layers that complement the output connected to the output layer of the network. The key idea is to improve the training process of a neural network through a variant of the standard backpropagation algorithm that considers these auxiliary outputs. The article presents experimental evidence of the advantages of the proposed idea in various real datasets. Results also show new research venues and practical applications into image recognition considering a deep learning setting.

Original languageEnglish
Title of host publicationPattern Recognition and Image Analysis - 9th Iberian Conference, IbPRIA 2019, Proceedings
EditorsAythami Morales, Julian Fierrez, José Salvador Sánchez, Bernardete Ribeiro
PublisherSpringer
Pages206-215
Number of pages10
ISBN (Print)9783030313319
DOIs
Publication statusPublished - 1 Jan 2019
Event9th Iberian Conference on Pattern Recognition and Image Analysis, IbPRIA 2019 - Madrid, Spain
Duration: 1 Jul 20194 Jul 2019

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11867 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference9th Iberian Conference on Pattern Recognition and Image Analysis, IbPRIA 2019
Country/TerritorySpain
CityMadrid
Period1/07/194/07/19

Keywords

  • Data classification
  • Neural network
  • Supervised learning

ASJC Scopus subject areas

  • Theoretical Computer Science
  • General Computer Science

Fingerprint

Dive into the research topics of 'A Proposal of Neural Networks with Intermediate Outputs'. Together they form a unique fingerprint.

Cite this