Title: SiNoptiC: swarm intelligence optimisation of convolutional neural network architectures for text classification
Authors: Imen Ferjani; Minyar Sassi Hidri; Ali Frihida
Addresses: Laboratory of Robotics, Informatics, and Complex Systems (RISC Lab - LR16ES07), National Engineering School of Tunis, University of Tunis El Manar, BP. 37, Le Belvedere, 1002, Tunis, Tunisia ' Computer Department, Deanship of Preparatory Year and Supporting Studies, Imam Abdulrahman Bin Faisal University, P.O. Box 1982, Dammam 31441, Saudi Arabia ' Laboratory of Robotics, Informatics, and Complex Systems (RISC Lab - LR16ES07), National Engineering School of Tunis, University of Tunis El Manar, BP. 37, Le Belvedere, 1002, Tunis, Tunisia
Abstract: Although many rules have been suggested by several researchers for designing deep neural architectures, trial-and-error is often exploited in practice to find the optimal model for a given problem. Thus, the automation of deep neural architecture search methods is highly recommended. In this work, we address this problem by proposing a hybrid coupling of Convolutional Neural Networks (CNNs) architectures with the swarm intelligence, especially the Fish School Search (FSS) algorithm. This coupling is capable of discovering a promising architecture of a CNN on handling text classification tasks. The proposed method allows users to provide training data as input, and receive a CNN model as an output. It is completely automatic and capable of fast convergence. Computational results show the effectiveness of the proposed method in achieving the best classification loss among manually designed CNNs. This is the first work using FSS for automatically designing the architectures of CNNs.
Keywords: deep learning; CNN; convolutional neural networks; swarm intelligence; FSS; text classification; NLP.
International Journal of Computer Applications in Technology, 2022 Vol.68 No.1, pp.82 - 100
Received: 18 Mar 2021
Accepted: 15 May 2021
Published online: 06 Jun 2022 *