Impact of Training Data Volume on Neural Network Training and Accuracy

Abstract

This master thesis explores the impact of data volume on model training and accuracy in the context of neural networks. The study focuses on conducting experiments on two image-based networks performing classification tasks, namely ResNet50 and MobileNetV2. The objective is to investigate the behaviour and accuracy of these networks as they are trained on progressively smaller subsets of the original dataset. With this study, we aim to gain some insight into how neural networks perform under different data availability scenarios. This type of information can become key in decision making processes regarding data collection, model development, and output handling, particularly in situations where data volume is limited. The research begins by establishing a baseline performance of the networks when trained on the entire dataset. Subsequently, various subsets of the original dataset are created by progressively reducing the volume of training data. The performance of the networks is then evaluated using these reduced datasets. This process allows for a comprehensive analysis of the effect of data volume on model training and accuracy. Throughout all of this process, statistical studies will be carried out to verify the robustness of our results, as well as the possible influence the different subsets have on the results. More specifically, the experiments involve training ResNet50 and MobileNetV2 models on subsets of the ImageNet-1K dataset, containing over 1.2 million training images across 1000 categories. The study examines how the reduction in training data volume affects the convergence of the models, as well as their accuracy in classifying images. Furthermore, the evolution of the network’s confidence in its predictions evolves through training.