Exploring Kervolutional Neural Networks
This is an incremental analysis of a previously proposed method, focusing on hyperparameter tuning and additional kervolution operations for computer vision researchers.
This paper analyzes kervolutional neural networks (KNNs), which use kervolution operations instead of standard convolution, finding they achieve faster convergence and higher accuracy than CNNs, with experiments showing up to 2.5% accuracy improvement on CIFAR-10.
A paper published in the CVPR 2019 conference outlines a new technique called 'kervolution' used in a new type of augmented convolutional neural network (CNN) called a 'kervolutional neural network' (KNN). The paper asserts that KNNs achieve faster convergence and higher accuracies than CNNs. This "mini paper" will further examine the findings in the original paper and perform a more in depth analysis of the KNN architecture. This will be done by analyzing the impact of hyper parameters (specifically the learning rate) on KNNs versus CNNs, experimenting with other types of kervolution operations not tested in the original paper, a more rigourous statistical analysis of accuracies and convergence times and additional theoretical analysis. The accompanying code is publicly available.