Abstract
Convolutional neural networks are generally assumed to be the best neural networks for classifying images. In November 2017, Geoffrey Hinton et al. introduced another neural network approach known as Capsule Network. He and his team have claimed that capsule networks are better than convolutional neural network, based on tests they performed on datasets such as MNIST and CIFAR-10. The aim of this project is to test capsule networks on distracted driver dataset and at the same time test the same dataset on convolutional network. This would help in comparing the performance of both types of neural networks. Geoffrey Hinton et al. believed that capsule networks can perform better than convolutional neural network when it comes to number of training images required to generalize the model. They also claimed that capsule networks have a better capability of classifying images keeping in mind the hierarchy between features of an image. This project would help in testing those claims. The goal of this project was also to create a network which would detect the distracted drivers which will help in reduction of accidents. After conducting experiments on both the networks, Capsule Network did train the model efficiently with a small number of training images. But, the accuracy was insufficient to say that we have a system in place to detect the distracted driver. When convolutional network was trained with a larger number of images, they gave a better accuracy than capsule networks. In addition, they were much faster than the capsule networks with respect to computation time.