Abstract:
Optimization is the process of adjusting the input parameters in a way to achieve the best possible result. The choice of hyper-parameters of a model varies from one dataset to another. Similarly, it is a difficult job to choose among different hyper-parameters available for a model. For example, the number of suitable layers, number of neurons, learning rate and activation function, etc. Moreover, it is difficult to determine the values of these parameters manually. The optimization of hyperparameters is thought to be particularly important for the accuracy of an algorithm. For parameter tuning, researchers use existing optimization algorithms or may also propose new optimization algorithms. In recent years, deep neural networks (DNN) have demonstrated good results for image classification and recognition. Convolutional neural network (CNN) is one of the most important and prominent models of DNN for image/video classification and pattern recognition. Various applications of CNN can be seen in medical image analysis, face recognition, lesion detection, speech recognition, and natural language processing. The model of CNN involves setting a number of parameters that may generate different results based on their configuration. Hyper-parameters play a crucial role in determining the accuracy and convergence of the CNN. The parameters of a model are often set through a random search or by adjusting manually using trial and error. In order to solve the problem of optimal parameters of a CNN, various researchers have proposed different evolutionary computation approaches to automatically design the optimal CNN architectures. Researchers have achieved better results when compared to state-of-the-art architectures. Different strategies for optimizing CNN architectures are proposed based on evolutionary computing techniques and these were evaluated on benchmark data sets and yielded competitive results. The effectiveness of Hunger Games search (HGS) with 10 state-of-the-art advanced algorithms has been illustrated with a few advanced meta-heuristic algorithms, different variants of DE on a comprehensive collection of 23 benchmark functions and IEEE CEC 2014 functions and some engineering problems. The proposed model is applied to the classification of CIFAR10 image dataset. CNNs have a large number of parameters and they can generate various classification accuracy for the same tasks based on diverse parameters including input of same scale and color to proceed. The HGS in this research is utilized to optimize the parameters of the CNN. Therefore, an initial CNN architecture is defined. During the optimization process of the HGS, the parametersare randomly initialized in the specified range, then recalculate and update to achieve the optimal CNN-HGS model. The experimental results obtained from this study demonstrate that the HGS algorithm outperforms other conventional optimization algorithms in terms of accuracy. HGS leads to superior performance in optimizing the Hyperparameter of CNNs for image classification. When compared to other commonly used optimization algorithms such as Simplified Swarm Optimization, particle swarm optimization, and differential evolution. The HGS algorithm consistently achieves higher accuracy rates on the benchmark CIFAR-10 dataset. According to the results, the proposed methodologies achieved favorable results with an accuracy of higher than 90%, showing competitive results compared to other algorithms. The superiority of the HGS algorithm in accuracy reaffirms its effectiveness in guiding the optimization process, resulting in more accurate and reliable CNN models for image classification tasks.
Page(s):
17-17
DOI:
DOI not available
Published:
Journal: Abstract Book on International Conference on Food and Applied Sciences (ICFAS-23) 3-5 August 23, Volume: 0, Issue: 0, Year: 2023
Keywords:
optimization
,
CNN
,
Classification
,
Hunger games search
,
CIFAR10