[Home ] [Archive]   [ فارسی ]  
:: Main :: About :: Current Issue :: Archive :: Search :: Submit :: Contact ::
Main Menu
Home::
Journal Information::
Articles archive::
For Authors::
For Reviewers::
Registration::
Contact us::
Site Facilities::
Publication Statistics::
List of Reviewers::
Social Networks::
::
Search in website

Advanced Search
..
Receive site information
Enter your Email in the following box to receive the site news and information.
..
:: Volume 10, Issue 1 (3-2023) ::
2023, 10(1): 60-83 Back to browse issues page
Knowledge Transfer of Convolutional Neural Networks via a Novel Evolutionary Fine-tuning Strategy (EFS) for Environmental Sound Classification
Hamed Riazati Seresht * , Karim Mohammadi
, h.riazati.s@gmail.com
Abstract:   (1336 Views)
Insufficient training data is one of the main challenges of utilizing deep Convolutional Neural Networks (CNNs) for Environmental Sound Classification (ESC). As a promising solution, Transfer Learning (TL) has addressed this issue by adapting a network pre-trained on a large-scale dataset to the target task. In this paper, we demonstrate that not all neurons/kernels of every layer in CNN networks are equally utilized to process the inputs of different classes, but there is a specific subgroups of neurons/kernels in every layer that play the key role in classification of every output class. Based on this observation and due to similarities that exist between feature spaces of some source and target classes, we propose to concentrate the fine-tuning process only on those neurons/kernels that do need changes and have the greatest impact on misclassifying target data. To identify these neurons/kernels, we pose a nested optimization problem for which we propose an effective evolutionary approach as solution.  Compared to the conventional fine-tuning approach, our proposed method achieves absolute improvements of about 1.9% and 2.3% in accuracy on ESC-50 and DCASE-17, respectively; remarkable improvements produced not by adding augmented data but with a more efficient utilization of knowledge stored in the pre-trained network. It is noteworthy that the computation time overhead of the proposed evolutionary method is rather small (about one third of the time required to train the model from scratch.
Article number: 3
Keywords: Deep learning, Convolutional Neural Networks, Transfer Learning, Genetic Algorithm, Environmental sound classification
Full-Text [PDF 10162 kb]   (407 Downloads)    
Type of Study: Research | Subject: Pattern Recognition
Received: 2023/01/4 | Accepted: 2023/02/7 | Published: 2023/12/12
Send email to the article author

Add your comments about this article
Your username or Email:

CAPTCHA


XML   Persian Abstract   Print


Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:

Riazati Seresht H, Mohammadi K. Knowledge Transfer of Convolutional Neural Networks via a Novel Evolutionary Fine-tuning Strategy (EFS) for Environmental Sound Classification. Nonlinear Systems in Electrical Engineering 2023; 10 (1) : 3
URL: http://journals.sut.ac.ir/jnsee/article-1-433-en.html


Rights and permissions
Creative Commons License This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
Volume 10, Issue 1 (3-2023) Back to browse issues page
سامانه های غیرخطی در مهندسی برق Journal of Nonlinear Systems in Electrical Engineering
نشریه سامانه‌های غیرخطی در مهندسی برق در خصوص اصول اخلاقی انتشار مقاله، از توصیه‌های «کمیته بین‌المللی اخلاق نشر» موسوم به COPE و «منشور و موازین اخلاق پژوهش» مصوب معاونت پژوهش و فناوری وزارت علوم، تحقیقات و فناوری تبعیت می‌کند.
Persian site map - English site map - Created in 0.11 seconds with 36 queries by YEKTAWEB 4642