Regularization of deep neural networks with spectral dropout

Salman Khan, Munawar HAYAT, Fatih Porikli

Research output: Contribution to journalArticle

5 Citations (Scopus)

Abstract

The big breakthrough on the ImageNet challenge in 2012 was partially due to the ‘Dropout’ technique used to avoid overfitting. Here, we introduce a new approach called ‘Spectral Dropout’ to improve the generalization ability of deep neural networks. We cast the proposed approach in the form of regular Convolutional Neural Network (CNN) weight layers using a decorrelation transform with fixed basis functions. Our spectral dropout method prevents overfitting by eliminating weak and ‘noisy’ Fourier domain coefficients of the neural network activations, leading to remarkably better results than the current regularization methods. Furthermore, the proposed is very efficient due to the fixed basis functions used for spectral transformation. In particular, compared to Dropout and Drop-Connect, our method significantly speeds up the network convergence rate during the training process (roughly ), with considerably higher neuron pruning rates (an increase of 30%). We demonstrate that the spectral dropout can also be used in conjunction with other regularization approaches resulting in additional performance gains.
Original languageEnglish
Pages (from-to)82-90
Number of pages9
JournalNeural Networks
Volume110
DOIs
Publication statusPublished - Feb 2019

Fingerprint

Neural networks
Neurons
Chemical activation
Weights and Measures
Deep neural networks

Cite this

Khan, Salman ; HAYAT, Munawar ; Porikli, Fatih. / Regularization of deep neural networks with spectral dropout. In: Neural Networks. 2019 ; Vol. 110. pp. 82-90.
@article{dddb4a4860694a1890524619f357f6f9,
title = "Regularization of deep neural networks with spectral dropout",
abstract = "The big breakthrough on the ImageNet challenge in 2012 was partially due to the ‘Dropout’ technique used to avoid overfitting. Here, we introduce a new approach called ‘Spectral Dropout’ to improve the generalization ability of deep neural networks. We cast the proposed approach in the form of regular Convolutional Neural Network (CNN) weight layers using a decorrelation transform with fixed basis functions. Our spectral dropout method prevents overfitting by eliminating weak and ‘noisy’ Fourier domain coefficients of the neural network activations, leading to remarkably better results than the current regularization methods. Furthermore, the proposed is very efficient due to the fixed basis functions used for spectral transformation. In particular, compared to Dropout and Drop-Connect, our method significantly speeds up the network convergence rate during the training process (roughly ), with considerably higher neuron pruning rates (an increase of 30{\%}). We demonstrate that the spectral dropout can also be used in conjunction with other regularization approaches resulting in additional performance gains.",
keywords = "Deep learning, Image classification, Regularization, Spectral analysis",
author = "Salman Khan and Munawar HAYAT and Fatih Porikli",
year = "2019",
month = "2",
doi = "10.1016/j.neunet.2018.09.009",
language = "English",
volume = "110",
pages = "82--90",
journal = "Neural Networks",
issn = "0893-6080",
publisher = "Elsevier Limited",

}

Regularization of deep neural networks with spectral dropout. / Khan, Salman; HAYAT, Munawar; Porikli, Fatih.

In: Neural Networks, Vol. 110, 02.2019, p. 82-90.

Research output: Contribution to journalArticle

TY - JOUR

T1 - Regularization of deep neural networks with spectral dropout

AU - Khan, Salman

AU - HAYAT, Munawar

AU - Porikli, Fatih

PY - 2019/2

Y1 - 2019/2

N2 - The big breakthrough on the ImageNet challenge in 2012 was partially due to the ‘Dropout’ technique used to avoid overfitting. Here, we introduce a new approach called ‘Spectral Dropout’ to improve the generalization ability of deep neural networks. We cast the proposed approach in the form of regular Convolutional Neural Network (CNN) weight layers using a decorrelation transform with fixed basis functions. Our spectral dropout method prevents overfitting by eliminating weak and ‘noisy’ Fourier domain coefficients of the neural network activations, leading to remarkably better results than the current regularization methods. Furthermore, the proposed is very efficient due to the fixed basis functions used for spectral transformation. In particular, compared to Dropout and Drop-Connect, our method significantly speeds up the network convergence rate during the training process (roughly ), with considerably higher neuron pruning rates (an increase of 30%). We demonstrate that the spectral dropout can also be used in conjunction with other regularization approaches resulting in additional performance gains.

AB - The big breakthrough on the ImageNet challenge in 2012 was partially due to the ‘Dropout’ technique used to avoid overfitting. Here, we introduce a new approach called ‘Spectral Dropout’ to improve the generalization ability of deep neural networks. We cast the proposed approach in the form of regular Convolutional Neural Network (CNN) weight layers using a decorrelation transform with fixed basis functions. Our spectral dropout method prevents overfitting by eliminating weak and ‘noisy’ Fourier domain coefficients of the neural network activations, leading to remarkably better results than the current regularization methods. Furthermore, the proposed is very efficient due to the fixed basis functions used for spectral transformation. In particular, compared to Dropout and Drop-Connect, our method significantly speeds up the network convergence rate during the training process (roughly ), with considerably higher neuron pruning rates (an increase of 30%). We demonstrate that the spectral dropout can also be used in conjunction with other regularization approaches resulting in additional performance gains.

KW - Deep learning

KW - Image classification

KW - Regularization

KW - Spectral analysis

UR - http://www.scopus.com/inward/record.url?scp=85057844267&partnerID=8YFLogxK

UR - http://www.mendeley.com/research/regularization-deep-neural-networks-spectral-dropout

U2 - 10.1016/j.neunet.2018.09.009

DO - 10.1016/j.neunet.2018.09.009

M3 - Article

VL - 110

SP - 82

EP - 90

JO - Neural Networks

JF - Neural Networks

SN - 0893-6080

ER -