Comparative Analysis of Convolutional and Capsule Networks on Decreasing Dataset Sizes: Insights for Real-World Applications

This paper presents a comparative analysis of the performance of convolutional and capsule neural networks on a dataset with decreasing sizes of training data. The study evaluates both normal and pre-trained models on a dataset ranging from 10 to 7500 training samples to investigate the efficacy of capsule networks in handling smaller datasets, a common challenge in real-world applications. Results indicate that convolutional networks outperform capsule networks on larger datasets, but as the dataset size decreases, capsule networks show improved performance and eventually surpass convolutional networks. But with very small sizes, capsule networks show problems with their accuracy. Pre-training surprisingly does not improve the performance of CNN, especially for smaller datasets, however, pretrained CapsNet showed some improvements for smaller datasets. These findings suggest that capsule networks are more effective than convolutional networks for datasets with a specific size, and pre-training further improves their performance for smaller sizes. The study offers insights into the practical applicability of capsule networks in real-world scenarios with limited dataset sizes, aiding in the selection of appropriate network architectures for diverse datasets and applications.