Overfitting Defeat with Dropout for Image Classification in Convolutional Neural Networks
Kukuh Nugroho, Hendrawan Hendrawan, Iskandar Iskandar · 2024
Data is a crucial part of machine learning, used to create models whose patterns are defined with specific algorithms. The primary purpose of the machine learning process is to generate the best prediction models. However, in general, models that trained well in the training phase differ when used in the testing phase. This phenomenon is called overfitting, a common problem in machine learning, mainly when developing image classification models using Convolutional Neural Networks (CNNs). The main task of the output layer in CNNs is to classify the truth labels from the data input images. One machine learning performance is defined by how well the models can predict the truth labels. Overfitting is one of the problems that can influence the model's performance. Dropout is one of the regularization techniques intended to solve this problem. Our purpose in this research is to know the relationship between the use of dropout and CNN's performance in image classification from the point of view of overfitting problems. The experimental results stated that this technique can alleviate overfitting by 35% with a learning rate value 0.001.