Efficient neural network structure design method based on layer-by-layer progressive mode

Abstract

The invention discloses an efficient neural network structure design method based on a layer-by-layer progressive mode. The objective of the invention is to solve the problem of huge time overhead of neural network structure search, so that efficient neural network structure search is realized by adopting a layer-by-layer progressive method, the network structure of the layer is searched from a single-layer neural network structure, all currently found layers are fused and trained, then the network structure of the next layer is searched, and the specified layer number is reached. According to the method, an excellent neural network model which enables the search time to be within an acceptable range of common computing power, and is high in model accuracy and short in reasoning time delay can be found.

Description

technical field [0001] The invention relates to a method for designing a deep neural network structure in the field of artificial intelligence, in particular to an efficient neural network structure search method. Background technique [0002] In recent years, deep learning technology has achieved great success in a large number of computational vision tasks, and deep neural structure is a key factor in determining performance. Research on fully automatic neural structure search methods has received more and more attention in recent years. [0003] An excellent network model must have a unique model structure, and the design of these model structures requires sufficient experience. In the industry, the commonly used model structure is the residual network (Res-18, Res-34), which is applied in almost all fields of machine learning, such as image classification, target detection, semantic segmentation, etc., and can even combine two The three-dimensional convolutional neural ...

AI Technical Summary

Problems solved by technology

Future methods may further improve search

Images