Showing 1–1 of 1 results for author: Sousa, E V

ConformalLayers: A non-linear sequential neural network with associative layers

Authors: Eduardo Vera Sousa, Leandro A. F. Fernandes, Cristina Nader Vasconcelos

Abstract: Convolutional Neural Networks (CNNs) have been widely applied. But as the CNNs grow, the number of arithmetic operations and memory footprint also increase. Furthermore, typical non-linear activation functions do not allow associativity of the operations encoded by consecutive layers, preventing the simplification of intermediate steps by combining them. We present a new activation function that a… ▽ More Convolutional Neural Networks (CNNs) have been widely applied. But as the CNNs grow, the number of arithmetic operations and memory footprint also increase. Furthermore, typical non-linear activation functions do not allow associativity of the operations encoded by consecutive layers, preventing the simplification of intermediate steps by combining them. We present a new activation function that allows associativity between sequential layers of CNNs. Even though our activation function is non-linear, it can be represented by a sequence of linear operations in the conformal model for Euclidean geometry. In this domain, operations like, but not limited to, convolution, average pooling, and dropout remain linear. We take advantage of associativity to combine all the "conformal layers" and make the cost of inference constant regardless of the depth of the network. △ Less

Submitted 9 November, 2021; v1 submitted 22 October, 2021; originally announced October 2021.

Comments: Best Paper on Pattern Recognition and Related Field at SIBGRAPI 2021 -- 34th Conference on Graphics, Patterns and Images