Determining the Minimal Number of Images Required to Effectively Train Convolutional Neural Networks

摘要

Determining the minimal number of images required to effectively train convolutional neural networks (CNNs) is not trivial and no simple solution exists. In this paper, we propose a method that helps to determine that number of images and checks if sufficient data is collected. We did a series of experiments using different CNN architectures (classic one with 3 convolutional layers, ResNet18, and Inception v3) and images data sets with different complexity (MNIST, CIFAR-10, CIFAR-100, and STL-10). These experiments allowed us to propose guidelines for the discussed problem. We also proposed a method for establishing the complexity of image sets.