2018/11/28 by Saeid Safaei, Safaei, Saeid, Vahid Safaei +9
Computer Science · #68Q32 #68T05 #Advanced Neural Network Applications #FOS: Computer and information sciences #I.2.6 #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Machine Learning and Data Classification #Neural Networks and Applications
paper · pdf · doi:10.48550/arxiv.1811.11813
openalex publication_date 2018/11/28 · openalex created_date 2022/08/01 · openalex updated_date 2026/07/28
The performance of artificial neural networks (ANNs) is influenced by weight\ninitialization, the nature of activation functions, and their architecture.\nThere is a wide range of activation functions that are traditionally used to\ntrain a neural network, e.g. sigmoid, tanh, and Rectified Linear Unit (ReLU). A\nwidespread practice is to use the same type of activation function in all\nneurons in a given layer. In this manuscript, we present a type of neural\nnetwork in which the activation functions in every layer form a polynomial\nbasis; we name this method SWAG after the initials of the last names of the\nauthors. We tested SWAG on three complex highly non-linear functions as well as\nthe MNIST handwriting data set. SWAG outperforms and converges faster than the\nstate of the art performance in fully connected neural networks. Given the low\ncomputational complexity of SWAG, and the fact that it was capable of solving\nproblems current architectures cannot, it has the potential to change the way\nthat we approach deep learning.\n