UFSCar/USP joint Seminar
Scheduled for:
May 7, 2021, at 2:00 pm
Speaker: Carlos Tadeu Pagani Zanini (Federal University of Rio de Janeiro)
Title: Radial Neural Networks
Abstract: This work proposes a very simple extension of the usual fully connected hidden layers in deep neural networks for classification. The objective is to transform the latent space on the hidden layers to be more suitable for the linear separation that occurs in the sigmoid/softmax output layer. We call such architectures radial neural networks because they use projections of fully connected hidden layers onto the surface of a hypersphere. We provide a geometrical motivation for the proposed method and show that it helps achieve convergence faster than the analogous architectures that they are built upon. As a result, we can significantly reduce training time on neural networks for classification that use fully connected hidden layers. The method is illustrated as an application to image classification, although it can be used for other classification tasks.
Bio:
Carlos Zanini is Assistant Professor at Federal University of Rio de Janeiro (UFRJ), Brazil, in the Department of Statistical Methods. He received his PhD (2019) from the Department of Statistics & Data Sciences at University of Texas (UT), Austin, USA. His B.A. and M.S. are from UFRJ. He is interested in Biostatistics, Bayesian non parametric methods, mixture models, dynamic models and neural networks.
