Artificial Bee Colony Algorithm Based Hyper-Parameter Optimization for Convolutional Neural Networks
Loading...
Date
2021-05-21
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Yalova University
Abstract
In recent years, Deep Learning has become a field that researchers are particularly interested in. The Convolutional Neural Network (CNN) is a type of multi-layered artificial neural network mostly used in the analysis, recognition, and classification of images and videos. The performance of CNN models is usually based on custom model architectures, thus several hyper-parameter values in a CNN are manually selected mostly. However, different combinations of hyper-parameters and models need to be used to achieve better performance results. Determination of optimum values of hyper parameters is also an optimization problem. The metaheuristic optimization techniques are able to solve such problems. In this paper, we propose to use the Artificial Bee Colony (ABC) algorithm which is a meta-heuristic approach to automatically determine the optimum architecture of a CNN by means of hyper-parameters. The most effective hyperparameters in the performance of CNN models have been optimized, which are the number of layers, the number and size of filters, activation function, batch size, learning rate, optimizer, and dropout rate. We have evaluated our optimized architecture using the well-known Fashion-MNIST dataset. The results demonstrate that the proposed model using ABC improves the performance of a CNN model.
Description
Keywords
convolutional neural networks, artificial bee colony, hyper-parameters, optimization, deep learning