Towards more efficient initialization methods for Convolutional Neural Networks via K-Means and Principal Components

This paper presents an exploration of unsupervised methods for initializing and training filters in convolutional layers, aiming to reduce the dependency on labeled data and computational resources. We propose two unsupervised methods based on the distribution of input data and evaluate their perfo...

Full description

Saved in:
Bibliographic Details
Main Authors: Federico Rabinovich, Facundo Quiroga, Franco Ronchetti
Format: Article
Language:English
Published: Postgraduate Office, School of Computer Science, Universidad Nacional de La Plata 2025-04-01
Series:Journal of Computer Science and Technology
Subjects:
Online Access:https://journal.info.unlp.edu.ar/JCST/article/view/3490
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850281043801669632
author Federico Rabinovich
Facundo Quiroga
Franco Ronchetti
author_facet Federico Rabinovich
Facundo Quiroga
Franco Ronchetti
author_sort Federico Rabinovich
collection DOAJ
description This paper presents an exploration of unsupervised methods for initializing and training filters in convolutional layers, aiming to reduce the dependency on labeled data and computational resources. We propose two unsupervised methods based on the distribution of input data and evaluate their performance against traditional Glorot Uniform initialization. By initializing solely the initial layer of a basic CNN network with one of our proposed methods, we attained a 0.78\% enhancement in final accuracy compared to traditional Glorot Uniform initialization. Our findings suggest that these unsupervised methods could serve as effective alternatives for filter initialization, potentially leading to more efficient training processes and a better understanding of CNNs.
format Article
id doaj-art-0b989766b58640418b00308484c68e5d
institution OA Journals
issn 1666-6046
1666-6038
language English
publishDate 2025-04-01
publisher Postgraduate Office, School of Computer Science, Universidad Nacional de La Plata
record_format Article
series Journal of Computer Science and Technology
spelling doaj-art-0b989766b58640418b00308484c68e5d2025-08-20T01:48:29ZengPostgraduate Office, School of Computer Science, Universidad Nacional de La PlataJournal of Computer Science and Technology1666-60461666-60382025-04-0125110.24215/16666038.25.e04Towards more efficient initialization methods for Convolutional Neural Networks via K-Means and Principal ComponentsFederico Rabinovich0Facundo Quiroga1Franco Ronchetti2UNLPIII-LIDI, Facultad de Informática, UNLP - CICIII-LIDI, Facultad de Informática, UNLP - CIC This paper presents an exploration of unsupervised methods for initializing and training filters in convolutional layers, aiming to reduce the dependency on labeled data and computational resources. We propose two unsupervised methods based on the distribution of input data and evaluate their performance against traditional Glorot Uniform initialization. By initializing solely the initial layer of a basic CNN network with one of our proposed methods, we attained a 0.78\% enhancement in final accuracy compared to traditional Glorot Uniform initialization. Our findings suggest that these unsupervised methods could serve as effective alternatives for filter initialization, potentially leading to more efficient training processes and a better understanding of CNNs. https://journal.info.unlp.edu.ar/JCST/article/view/3490ClusteringCNN, Deep learning,K-Means ClusteringPrincipal Component Analysis
spellingShingle Federico Rabinovich
Facundo Quiroga
Franco Ronchetti
Towards more efficient initialization methods for Convolutional Neural Networks via K-Means and Principal Components
Journal of Computer Science and Technology
Clustering
CNN, Deep learning,
K-Means Clustering
Principal Component Analysis
title Towards more efficient initialization methods for Convolutional Neural Networks via K-Means and Principal Components
title_full Towards more efficient initialization methods for Convolutional Neural Networks via K-Means and Principal Components
title_fullStr Towards more efficient initialization methods for Convolutional Neural Networks via K-Means and Principal Components
title_full_unstemmed Towards more efficient initialization methods for Convolutional Neural Networks via K-Means and Principal Components
title_short Towards more efficient initialization methods for Convolutional Neural Networks via K-Means and Principal Components
title_sort towards more efficient initialization methods for convolutional neural networks via k means and principal components
topic Clustering
CNN, Deep learning,
K-Means Clustering
Principal Component Analysis
url https://journal.info.unlp.edu.ar/JCST/article/view/3490
work_keys_str_mv AT federicorabinovich towardsmoreefficientinitializationmethodsforconvolutionalneuralnetworksviakmeansandprincipalcomponents
AT facundoquiroga towardsmoreefficientinitializationmethodsforconvolutionalneuralnetworksviakmeansandprincipalcomponents
AT francoronchetti towardsmoreefficientinitializationmethodsforconvolutionalneuralnetworksviakmeansandprincipalcomponents