Harmonic Networks with Limited Training Samples. Ulicny, M., Krylov, V. A., & Dahyot, R. In 2019 27th European Signal Processing Conference (EUSIPCO), pages 1-5, Sep., 2019. Github: https://github.com/matej-ulicny/harmonic-networks and paper also on arxiv http://arxiv.org/abs/1905.00135 and https://www.eurasip.org/Proceedings/Eusipco/eusipco2019/Proceedings/papers/1570533913.pdfPaper doi abstract bibtex Convolutional neural networks (CNNs) are very popular nowadays for image processing. CNNs allow one to learn optimal filters in a (mostly) supervised machine learning context. However this typically requires abundant labelled training data to estimate the filter parameters. Alternative strategies have been deployed for reducing the number of parameters and / or filters to be learned and thus decrease overfitting. In the context of reverting to preset filters, we propose here a computationally efficient harmonic block that uses Discrete Cosine Transform (DCT) filters in CNNs. In this work we examine the performance of harmonic networks in limited training data scenario. We validate experimentally that its performance compares well against scattering networks that use wavelets as preset filters.
@INPROCEEDINGS{8902831,
author = {M. {Ulicny} and V. A. {Krylov} and R. {Dahyot}},
booktitle = {2019 27th European Signal Processing Conference (EUSIPCO)},
url = {https://mural.maynoothuniversity.ie/15158/1/RD_harmonic%20networks.pdf},
title = {Harmonic Networks with Limited Training Samples},
year = {2019},
volume = {},
number = {},
pages = {1-5},
abstract = {Convolutional neural networks (CNNs) are very popular nowadays for image processing. CNNs allow one to learn optimal filters in a (mostly) supervised machine learning context. However this typically requires abundant labelled training data to estimate the filter parameters. Alternative strategies have been deployed for reducing the number of parameters and / or filters to be learned and thus decrease overfitting. In the context of reverting to preset filters, we propose here a computationally efficient harmonic block that uses Discrete Cosine Transform (DCT) filters in CNNs. In this work we examine the performance of harmonic networks in limited training data scenario. We validate experimentally that its performance compares well against scattering networks that use wavelets as preset filters.},
keywords = {Lapped Discrete Cosine Transform;harmonic network;convolutional filter;limited data},
doi = {10.23919/EUSIPCO.2019.8902831},
note = {Github: https://github.com/matej-ulicny/harmonic-networks and paper also on arxiv http://arxiv.org/abs/1905.00135 and https://www.eurasip.org/Proceedings/Eusipco/eusipco2019/Proceedings/papers/1570533913.pdf},
archivePrefix = {arXiv},
eprint = {1905.00135},
ISSN = {2219-5491},
month = {Sep.}}
Downloads: 0
{"_id":"nLsJ7GTrNyiPdPsmh","bibbaseid":"ulicny-krylov-dahyot-harmonicnetworkswithlimitedtrainingsamples-2019","authorIDs":["6fptrFgK7WSZkf6TM"],"author_short":["Ulicny, M.","Krylov, V. A.","Dahyot, R."],"bibdata":{"bibtype":"inproceedings","type":"inproceedings","author":[{"firstnames":["M."],"propositions":[],"lastnames":["Ulicny"],"suffixes":[]},{"firstnames":["V.","A."],"propositions":[],"lastnames":["Krylov"],"suffixes":[]},{"firstnames":["R."],"propositions":[],"lastnames":["Dahyot"],"suffixes":[]}],"booktitle":"2019 27th European Signal Processing Conference (EUSIPCO)","url":"https://mural.maynoothuniversity.ie/15158/1/RD_harmonic%20networks.pdf","title":"Harmonic Networks with Limited Training Samples","year":"2019","volume":"","number":"","pages":"1-5","abstract":"Convolutional neural networks (CNNs) are very popular nowadays for image processing. CNNs allow one to learn optimal filters in a (mostly) supervised machine learning context. However this typically requires abundant labelled training data to estimate the filter parameters. Alternative strategies have been deployed for reducing the number of parameters and / or filters to be learned and thus decrease overfitting. In the context of reverting to preset filters, we propose here a computationally efficient harmonic block that uses Discrete Cosine Transform (DCT) filters in CNNs. In this work we examine the performance of harmonic networks in limited training data scenario. We validate experimentally that its performance compares well against scattering networks that use wavelets as preset filters.","keywords":"Lapped Discrete Cosine Transform;harmonic network;convolutional filter;limited data","doi":"10.23919/EUSIPCO.2019.8902831","note":"Github: https://github.com/matej-ulicny/harmonic-networks and paper also on arxiv http://arxiv.org/abs/1905.00135 and https://www.eurasip.org/Proceedings/Eusipco/eusipco2019/Proceedings/papers/1570533913.pdf","archiveprefix":"arXiv","eprint":"1905.00135","issn":"2219-5491","month":"Sep.","bibtex":"@INPROCEEDINGS{8902831, \nauthor = {M. {Ulicny} and V. A. {Krylov} and R. {Dahyot}}, \nbooktitle = {2019 27th European Signal Processing Conference (EUSIPCO)},\nurl = {https://mural.maynoothuniversity.ie/15158/1/RD_harmonic%20networks.pdf},\ntitle = {Harmonic Networks with Limited Training Samples}, \nyear = {2019}, \nvolume = {}, \nnumber = {}, \npages = {1-5}, \nabstract = {Convolutional neural networks (CNNs) are very popular nowadays for image processing. CNNs allow one to learn optimal filters in a (mostly) supervised machine learning context. However this typically requires abundant labelled training data to estimate the filter parameters. Alternative strategies have been deployed for reducing the number of parameters and / or filters to be learned and thus decrease overfitting. In the context of reverting to preset filters, we propose here a computationally efficient harmonic block that uses Discrete Cosine Transform (DCT) filters in CNNs. In this work we examine the performance of harmonic networks in limited training data scenario. We validate experimentally that its performance compares well against scattering networks that use wavelets as preset filters.},\nkeywords = {Lapped Discrete Cosine Transform;harmonic network;convolutional filter;limited data}, \ndoi = {10.23919/EUSIPCO.2019.8902831},\nnote = {Github: https://github.com/matej-ulicny/harmonic-networks and paper also on arxiv http://arxiv.org/abs/1905.00135 and https://www.eurasip.org/Proceedings/Eusipco/eusipco2019/Proceedings/papers/1570533913.pdf},\narchivePrefix = {arXiv},\neprint = {1905.00135},\nISSN = {2219-5491}, \nmonth = {Sep.}}\n\n\n","author_short":["Ulicny, M.","Krylov, V. A.","Dahyot, R."],"key":"8902831","id":"8902831","bibbaseid":"ulicny-krylov-dahyot-harmonicnetworkswithlimitedtrainingsamples-2019","role":"author","urls":{"Paper":"https://mural.maynoothuniversity.ie/15158/1/RD_harmonic%20networks.pdf"},"keyword":["Lapped Discrete Cosine Transform;harmonic network;convolutional filter;limited data"],"metadata":{"authorlinks":{"dahyot, r":"https://roznn.github.io/ipublication.html"}},"downloads":0},"bibtype":"inproceedings","biburl":"https://raw.githubusercontent.com/Roznn/Roznn.github.io/master/works.bib","creationDate":"2021-01-17T18:19:29.523Z","downloads":0,"keywords":["lapped discrete cosine transform;harmonic network;convolutional filter;limited data"],"search_terms":["harmonic","networks","limited","training","samples","ulicny","krylov","dahyot"],"title":"Harmonic Networks with Limited Training Samples","year":2019,"dataSources":["LyGNQYGRFw8k9gPrK","NqWTiMfRR56v86wRs","dtJ7afty6nTMHhqAE","r6oz3cMyC99QfiuHW"]}