Solving multi-objective optimization problem of convolutional neural network using fast forward quantum optimization algorithm: Application in digital image classification. (February 2023)

• Record Type: Journal Article
• Authors: Singh, Pritpal; Muchahari, Monoj Kumar
• Is Part Of: Advances in engineering software. Volume 176(2023)
• Publication Date: 2023-02
• Subjects: Multi-objective Optimization Problem (MOOP) -- Convolutional neural networks (CNNs) -- Fast forward quantum optimization algorithm (FFQOA) -- Image classification; Computer-aided engineering -- Periodicals; Engineering -- Computer programs -- Periodicals; Engineering -- Software -- Periodicals; Periodicals; 620.0028553
• DOI: 10.1016/j.advengsoft.2022.103370
• Languages: English
• ISSNs: 0965-9978

Abstract

Abstract: Convolutional neural network (CNN) has evolved as a new algorithm that has demonstrated its effectiveness in real-time issue solving over many other machine learning (ML) algorithms. However, many CNNs are created manually by considering randomly defined weights in the convolutional layer, pooling layer, and fully connected layer. During the training process, these weights may get stuck at the local minima. To avoid this, the weights must be initialized carefully in each of the layers. Furthermore, training cannot be accomplished until the classification error is minimized. The classification error largely depends on the optimal values of the connecting weights among the layers. Since optimization of weights requires many iterations, it increases the difficulty of selecting the minimum value of classification error. Therefore, weights and classification errors are considered to be two critical parameters that influence the performance of CNNs, and they must be carefully governed during architecture design. However, finding the optimal values for weights and classification error can be considered a multi-objective optimization problem (MOOP). To solve a MOOP, this study advocates the use of the recently proposed fast forward quantum optimization algorithm (FFQOA). This study also proposes a novel algorithm based on the hybridization of FFQOA with CNN, called the FFQOAconNetwork. In this algorithm, the FFQOA searches for the optimal weights associated with the layers by simultaneously achieving the minimal classification error. Application of the FFQOAconNetwork is demonstrated in the classification of images by adopting benchmark datasets. Empirical analyses indicate that the FFQOAconNetwork can solve the MOOP with effective performance as compared to other algorithms. Highlights: This paper discusses the MOOP of convolutional neural network (CNN). To solve this MOOP, fast forward quantum optimization algorithm (FFQOA) is used. The FFQOA is hybridized with CNN, which is called FFQOAconNetwork. The FFQOAconNetwork is applied in the classification of benchmark digital images. The performance evaluation metrics show the effectiveness of FFQOAconNetwork.