CapsNet on Embedded Devices in a Data Scarce Scenario

This paper explores the use of the Capsule Neural Network (CapsNet) on small datasets for image recognition and its possible embedded/edge device implementation. As compared to data-hungry Convolutional Neural Network (CNN), CapsNet can operate with data scarce scenarios. CapsNet is briefly presented, followed by the training process. A CapsNet model is designed and trained, exploring the effects of various hyperparameters on its performance. With a two-class dataset of 30 images per class, our proposed CapsNet model yields training and validation accuracies of 70% after training on 35 epochs, but begins to overfit if trained beyond. An equivalent baseline CNN model already overfits before 35 epochs, with training and validation accuracies of 99% and 66.70%, respectively. This shows that with small datasets, a CNN overfits rapidly regardless of the number of epochs, while we can design a CapsN et model that performs considerably better on small datasets, if properly tuned. We also show that a CapsNet model can execute onto an embedded/edge device after compression for reducing its size and complexity.