Learning sparse filter bank transforms with convolutional ICA

Johannes Ballé, Eero Simoncelli

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Independent Component Analysis (ICA) is a generalization of Principal Component Analysis that optimizes a linear transformation to whiten and sparsify a family of source signals. The computational costs of ICA grow rapidly with dimensionality, and application to high-dimensional data is generally achieved by restricting to small windows, violating the translation-invariant nature of many real-world signals, and producing blocking artifacts in applications. Here, we reformulate the ICA problem for transformations computed through convolution with a bank of filters, and develop a generalization of the fastICA algorithm for optimizing the filters over a set of example signals. This results in a substantial reduction of computational complexity and memory requirements. When applied to a database of photographic images, the method yields bandpass oriented filters, whose responses are sparser than those of orthogonal wavelets or block DCT, and slightly more heavy-tailed than those of block ICA, despite fewer model parameters.

Original languageEnglish (US)
Title of host publication2014 IEEE International Conference on Image Processing, ICIP 2014
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages4013-4017
Number of pages5
ISBN (Print)9781479957514
DOIs
StatePublished - Jan 28 2014

Fingerprint

Filter banks
Independent component analysis
Linear transformations
Convolution
Principal component analysis
Computational complexity
Data storage equipment
Costs

Keywords

  • convolutional filters
  • fastICA
  • filter bank
  • independent component analysis
  • sparsity
  • stationarity

ASJC Scopus subject areas

  • Computer Vision and Pattern Recognition

Cite this

Ballé, J., & Simoncelli, E. (2014). Learning sparse filter bank transforms with convolutional ICA. In 2014 IEEE International Conference on Image Processing, ICIP 2014 (pp. 4013-4017). [7025815] Institute of Electrical and Electronics Engineers Inc.. https://doi.org/10.1109/ICIP.2014.7025815

Learning sparse filter bank transforms with convolutional ICA. / Ballé, Johannes; Simoncelli, Eero.

2014 IEEE International Conference on Image Processing, ICIP 2014. Institute of Electrical and Electronics Engineers Inc., 2014. p. 4013-4017 7025815.

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Ballé, J & Simoncelli, E 2014, Learning sparse filter bank transforms with convolutional ICA. in 2014 IEEE International Conference on Image Processing, ICIP 2014., 7025815, Institute of Electrical and Electronics Engineers Inc., pp. 4013-4017. https://doi.org/10.1109/ICIP.2014.7025815
Ballé J, Simoncelli E. Learning sparse filter bank transforms with convolutional ICA. In 2014 IEEE International Conference on Image Processing, ICIP 2014. Institute of Electrical and Electronics Engineers Inc. 2014. p. 4013-4017. 7025815 https://doi.org/10.1109/ICIP.2014.7025815
Ballé, Johannes ; Simoncelli, Eero. / Learning sparse filter bank transforms with convolutional ICA. 2014 IEEE International Conference on Image Processing, ICIP 2014. Institute of Electrical and Electronics Engineers Inc., 2014. pp. 4013-4017
@inproceedings{6b326002e58148f3b3cd1d548b901f25,
title = "Learning sparse filter bank transforms with convolutional ICA",
abstract = "Independent Component Analysis (ICA) is a generalization of Principal Component Analysis that optimizes a linear transformation to whiten and sparsify a family of source signals. The computational costs of ICA grow rapidly with dimensionality, and application to high-dimensional data is generally achieved by restricting to small windows, violating the translation-invariant nature of many real-world signals, and producing blocking artifacts in applications. Here, we reformulate the ICA problem for transformations computed through convolution with a bank of filters, and develop a generalization of the fastICA algorithm for optimizing the filters over a set of example signals. This results in a substantial reduction of computational complexity and memory requirements. When applied to a database of photographic images, the method yields bandpass oriented filters, whose responses are sparser than those of orthogonal wavelets or block DCT, and slightly more heavy-tailed than those of block ICA, despite fewer model parameters.",
keywords = "convolutional filters, fastICA, filter bank, independent component analysis, sparsity, stationarity",
author = "Johannes Ball{\'e} and Eero Simoncelli",
year = "2014",
month = "1",
day = "28",
doi = "10.1109/ICIP.2014.7025815",
language = "English (US)",
isbn = "9781479957514",
pages = "4013--4017",
booktitle = "2014 IEEE International Conference on Image Processing, ICIP 2014",
publisher = "Institute of Electrical and Electronics Engineers Inc.",

}

TY - GEN

T1 - Learning sparse filter bank transforms with convolutional ICA

AU - Ballé, Johannes

AU - Simoncelli, Eero

PY - 2014/1/28

Y1 - 2014/1/28

N2 - Independent Component Analysis (ICA) is a generalization of Principal Component Analysis that optimizes a linear transformation to whiten and sparsify a family of source signals. The computational costs of ICA grow rapidly with dimensionality, and application to high-dimensional data is generally achieved by restricting to small windows, violating the translation-invariant nature of many real-world signals, and producing blocking artifacts in applications. Here, we reformulate the ICA problem for transformations computed through convolution with a bank of filters, and develop a generalization of the fastICA algorithm for optimizing the filters over a set of example signals. This results in a substantial reduction of computational complexity and memory requirements. When applied to a database of photographic images, the method yields bandpass oriented filters, whose responses are sparser than those of orthogonal wavelets or block DCT, and slightly more heavy-tailed than those of block ICA, despite fewer model parameters.

AB - Independent Component Analysis (ICA) is a generalization of Principal Component Analysis that optimizes a linear transformation to whiten and sparsify a family of source signals. The computational costs of ICA grow rapidly with dimensionality, and application to high-dimensional data is generally achieved by restricting to small windows, violating the translation-invariant nature of many real-world signals, and producing blocking artifacts in applications. Here, we reformulate the ICA problem for transformations computed through convolution with a bank of filters, and develop a generalization of the fastICA algorithm for optimizing the filters over a set of example signals. This results in a substantial reduction of computational complexity and memory requirements. When applied to a database of photographic images, the method yields bandpass oriented filters, whose responses are sparser than those of orthogonal wavelets or block DCT, and slightly more heavy-tailed than those of block ICA, despite fewer model parameters.

KW - convolutional filters

KW - fastICA

KW - filter bank

KW - independent component analysis

KW - sparsity

KW - stationarity

UR - http://www.scopus.com/inward/record.url?scp=84949928273&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84949928273&partnerID=8YFLogxK

U2 - 10.1109/ICIP.2014.7025815

DO - 10.1109/ICIP.2014.7025815

M3 - Conference contribution

AN - SCOPUS:84949928273

SN - 9781479957514

SP - 4013

EP - 4017

BT - 2014 IEEE International Conference on Image Processing, ICIP 2014

PB - Institute of Electrical and Electronics Engineers Inc.

ER -