A Proposal of Neural Networks with Intermediate Outputs

Billy Peralta, Juan Reyes, Luis Caro, Christian Pieringer

Resultado de la investigación: Contribución a los tipos de informe/libroContribución a la conferenciarevisión exhaustiva

Resumen

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.

Idioma originalInglés
Título de la publicación alojadaPattern Recognition and Image Analysis - 9th Iberian Conference, IbPRIA 2019, Proceedings
EditoresAythami Morales, Julian Fierrez, José Salvador Sánchez, Bernardete Ribeiro
EditorialSpringer
Páginas206-215
Número de páginas10
ISBN (versión impresa)9783030313319
DOI
EstadoPublicada - 1 ene. 2019
Evento9th Iberian Conference on Pattern Recognition and Image Analysis, IbPRIA 2019 - Madrid, Espana
Duración: 1 jul. 20194 jul. 2019

Serie de la publicación

NombreLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volumen11867 LNCS
ISSN (versión impresa)0302-9743
ISSN (versión digital)1611-3349

Conferencia

Conferencia9th Iberian Conference on Pattern Recognition and Image Analysis, IbPRIA 2019
País/TerritorioEspana
CiudadMadrid
Período1/07/194/07/19

Áreas temáticas de ASJC Scopus

  • Ciencia computacional teórica
  • Informática (todo)

Huella

Profundice en los temas de investigación de 'A Proposal of Neural Networks with Intermediate Outputs'. En conjunto forman una huella única.

Citar esto