An alternative for kernel SVM when stacked with a neural network

Abstract

Many studies stack SVM and neural network by utilzing SVM as an output layer of the neural network. However, those studies use kernel before the SVM which is unnecessary. In this study, we proposed an alternative to kernel SVM and proved why kernel is unnecessary when the SVM is stacked on top of neural network. The experiments is done on Dublin City LiDAR data. In this study, we stack PointNet and SVM but instead of using kernel, we simply utilize the last hidden layer of the PointNet. As an alternative to the SVM kernel, this study performs dimension expansion by increasing the number of neurons in the last hidden layer. We proved that expanding the dimension by increasing the number of neurons in the last hidden layer can increase the F-Measure score and it performs better than RBF kernel both in term of F-Measure score and computation time.