EFFICIENT SIMILARITY-BASED PASSIVE FILTER PRUNING FOR COMPRESSING CNNS Arshdeep Singh Mark D. Plumbley

2025-08-25 0 0 826.54KB 5 页 10玖币

侵权投诉

EFFICIENT SIMILARITY-BASED PASSIVE FILTER PRUNING FOR COMPRESSING CNNS

Arshdeep Singh, Mark D. Plumbley

Centre for Vision, Speech and Signal Processing (CVSSP)

University of Surrey, UK

Email: {arshdeep.singh, m.plumbley}@surrey.ac.uk

ABSTRACT

Convolution neural networks (CNNs) have shown great success

in various applications. However, the computational complexity

and memory storage of CNNs is a bottleneck for their deployment

on resource-constrained devices. Recent efforts towards reducing

the computation cost and the memory overhead of CNNs involve

similarity-based passive ﬁlter pruning methods. Similarity-based

passive ﬁlter pruning methods compute a pairwise similarity ma-

trix for the ﬁlters and eliminate a few similar ﬁlters to obtain a

small pruned CNN. However, the computational complexity of

computing the pairwise similarity matrix is high, particularly when

a convolutional layer has many ﬁlters. To reduce the computational

complexity in obtaining the pairwise similarity matrix, we propose

to use an efﬁcient method where the complete pairwise similarity

matrix is approximated from only a few of its columns by using a

Nystr¨

om approximation method. The proposed efﬁcient similarity-

based passive ﬁlter pruning method is 3 times faster and gives same

accuracy at the same reduction in computations for CNNs com-

pared to that of the similarity-based pruning method that computes

a complete pairwise similarity matrix. Apart from this, the proposed

efﬁcient similarity-based pruning method performs similarly or bet-

ter than the existing norm-based pruning methods. The efﬁcacy of

the proposed pruning method is evaluated on CNNs such as DCASE

2021 Task 1A baseline network and a VGGish network designed for

acoustic scene classiﬁcation.

Index Terms—Acoustic scene classiﬁcation, pruning, VGGish,

DCASE.

1. INTRODUCTION

Compressing convolutional neural networks (CNNs) is crucial to re-

duce their computational complexity and memory storage for efﬁ-

cient deployment on resource-constrained devices [1], despite state-

of-the-art performances of CNNs in various applications [2]. Typ-

ically, CNNs have redundant parameters such as weights or ﬁlters,

which yield only extra computations and storage without contribut-

ing much to the performance of the underlying task [3, 4]. For exam-

ple, Singh et al. [5, 6] found that 73% of the ﬁlters in SoundNet that

do not provide discriminative information across different acoustic

scene classes, and eliminating such ﬁlters gives similar performance

compared to that of using all ﬁlters in SoundNet. Thus, the com-

pression of CNNs has recently drawn signiﬁcant attention from the

research community.

Recent efforts towards compressing CNNs involve ﬁlter pruning

methods [7, 8] that eliminate some of the ﬁlters in CNNs based on

their importance. The importance of the CNN ﬁlters is measured in

an active or in a passive manner. Active ﬁlter pruning methods in-

volve a dataset. For example, some methods [9, 10, 11] use feature

0.2 0.6

0 0.1

0 2

10.1

0 2

Fig. 1. An illustration of output produced in a convolution layer by

three CNN ﬁlters, F1,F2and F3, with a convolution operation on

randomly generated data points, X∈R2×1000.

maps which are outputs produced by the ﬁlters corresponding to a

set of examples, and apply metrics such as entropy or the average

percentage of zeros on the feature maps to quantify the ﬁlter impor-

tance. On the other hand, passive ﬁlter pruning methods [12, 13]

use only parameters of the ﬁlters, such as an absolute sum of the

weights in the ﬁlters, to quantify the ﬁlter importance. The pas-

sive ﬁlter pruning methods do not involve a dataset to measure ﬁlter

importance and therefore are easier to apply compared to active ﬁl-

ter pruning methods. After eliminating ﬁlters from the CNNs, the

pruned network is ﬁne-tuned to regain some of the performance lost

due to the ﬁlter elimination.

Previously, passive ﬁlter pruning methods used norm-based met-

rics such as l1-norm [12], which is a sum of the absolute values of

each weight in the ﬁlter, or l2-distance of the ﬁlters from a geomet-

ric median of all ﬁlters [13] to quantify the importance of the ﬁlters.

These norm-based methods use a “smaller-norm-less-important” cri-

terion to eliminate ﬁlters. For example, a ﬁlter having a relatively

high l1-norm is considered more important than others. However,

while selecting relatively high-norm ﬁlters as important, norm-based

methods may ignore the redundancy among the high-norm ﬁlters. To

illustrate this, we show outputs produced by three ﬁlters in Figure 1.

Filters F1and F3have similar l1-norm and produce similar outputs.

However, selecting two important ﬁlters out of the three ﬁlters shown

in Figure 1, the norm-based method selects ﬁlters F1and F3as im-

portant due to their relatively high norm, despite producing similar

outputs, while it eliminates ﬁlter F2that produces signiﬁcantly dif-

ferent output than the other ﬁlters. Thus the diversity learned in the

arXiv:2210.17416v1 [cs.CV] 27 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

EFFICIENTSIMILARITY-BASEDPASSIVEFILTERPRUNINGFORCOMPRESSINGCNNSArshdeepSingh,MarkD.PlumbleyCentreforVision,SpeechandSignalProcessing(CVSSP)UniversityofSurrey,UKEmail:farshdeep.singh,m.plumbleyg@surrey.ac.ukABSTRACTConvolutionneuralnetworks(CNNs)haveshowngreatsuccessinvariousapplications.However,thec...

展开>> 收起<<

EFFICIENT SIMILARITY-BASED PASSIVE FILTER PRUNING FOR COMPRESSING CNNS Arshdeep Singh Mark D. Plumbley.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

EFFICIENT SIMILARITY-BASED PASSIVE FILTER PRUNING FOR COMPRESSING CNNS Arshdeep Singh Mark D. Plumbley

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: