Multicriteria Optimization for Structural Parametric Synthesis of Convolutional Neural Networks

Abstract

The paper defines a promising class of convolutional neural networks and considers their key parameters for further structural and parametric synthesis. It is shown that these networks should include, in addition to traditional components (convolutional layers, pooling layers, feed-forward layers, additional layers: batch normalization layer, 1x1 convolutional layer, dropout layer, etc), also functional structural units (SRU, CRU, dense residual attention unit, etc). We propose to use a genetic algorithm for structural-parametric synthesis using the considered layers and structural blocks.