Automatically Designing CNN Architectures for Acoustic Scene Classification

• Published in: 2021 16th International Conference on Computer Engineering and Systems (ICCES) pp. 1 - 6
• Main Authors: Hasan, Noha W., Saudi, Ali S., Khalil, Mahmoud I., Abbas, Hazem M.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 15.12.2021
• Subjects: Acoustic scene classification (ASC); Acoustics; Classification algorithms; Computer architecture; Convolutional neural networks; convolutional neural networks (CNNs); genetic algorithms (GAs); Graphics processing units; Image analysis; Neural networks; neural-network architecture optimization
• DOI: 10.1109/ICCES54031.2021.9686085

Deep convolutional neural networks (CNNs) have proven their effectiveness in the acoustic scene classification (ASC) task, becoming the state-of-the-art solution to build pow-erful ASC models. Usually, the architectures of these networks and their hyper-parameters have been manually designed by expert researchers in both the investigated problem and CNN architecture design, which can be a challenging task. In this work, we propose a genetic algorithm (GA) that benefits from the characteristics of state-of-the-art CNN architectures in the field of ASC to specifically optimize CNN architectures for the ASC task, within an acceptable search cost compared to other GAs. After searching for 14 GPU days using a single NVIDIA Tesla K80 GPU on the development dataset of DCASE2020 Task lA, the proposed GA has optimized a CNN architecture that achieved a test accuracy of 68.5 % on the same dataset. The same network has been evaluated on the development dataset of DCASE2018 Task 1A and achieved a classification accuracy of 73.7% without repeating the search process, which demonstrates the generalization capability of the generated CNNs.