Peralta, Billy; Reyes, Juan; Caro, Luis; Pieringer, Christian

A Proposal of Neural Networks with Intermediate Outputs

datacite.creator	Peralta, Billy
datacite.creator	Reyes, Juan
datacite.creator	Caro, Luis
datacite.creator	Pieringer, Christian
datacite.date	2019-06-04
datacite.rights	Acceso Abierto
datacite.title	A Proposal of Neural Networks with Intermediate Outputs
dc.date.accessioned	2024-11-21T14:47:35Z
dc.date.available	2024-11-21T14:47:35Z
dc.description	Financiado por la Universidad Andrés Bello.
dc.description.abstracten	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.
dc.identifier.doi	10.1007/978-3-030-31332-6_18
dc.identifier.uri	https://repositoriodigital.uct.cl/handle/10925/6042
dc.language.iso	en
dc.publisher	Springer
dc.rights	Revista de acceso abierto registrada en el Directory of Open Access Journals (DOAJ) o en el Directory of Open Access Scholarly Resources (ROAD).
dc.source	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
oaire.citationEndPage	215
oaire.citationStartPage	206
oaire.citationTitle	Artículo
oaire.citationVolume	11867
oaire.resourceType	Artículo
uct.catalogador	mlj
uct.departamento	Departamento Ciencias Matematicas y Fisicas
uct.facultad	Facultad de Ingeniería
uct.indizacion	SCOPUS