الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 145 |  |  |
| | | from | 145 | | |
Abstract
This thesis proposes two different methods to automatically design CNN architectures for the acoustic scene classification task. The first method uses the differentiable ar- chitecture search approach to design appropriate CNN architectures for the task with a limited search cost of only a few GPU days. This method is based on the continuous re- laxation of the architecture representation that allows efficient search of the architecture space using gradient descent. This method has achieved a competitve classification accu- racy of 64.57±0.87% on the acoustic scene classification dataset of DCASE2020-Task1A. The second method is a novel genetic algorithm that benefits from the characteristics of state-of-the-art CNN architectures in the field of acoustic scene classification to specif- ically optimize CNN architectures for the task, within an acceptable search cost com- pared to other genetic algorithms. The proposed algorithm uses frequency-dimension splitting of the input spectrograms in order to explore the architecture search space in sub-CNN models in addition to classical single-path CNNs. Specifically, this algorithm aims to find the best number of sub-CNNs in addition to their architectures to better capture the distinct features of the input spectrograms. Experimental results on three benchmark datasets demonstrate the effectiveness of the proposed method. Specifically, the proposed algorithm has achieved around 17.8%, 16%, and 17.2%, relative improve- ment in accuracy with respect to the baseline systems on the development datasets of DCASE2018-Task1A, DCASE2019-Task1A, and DCASE2020-Task1A, respectively.