A structure searching method and device of a depth neural network

Abstract

The invention provides a structure searching method and device of a depth neural network, which relates to the technical field of artificial intelligence. The method comprises the following steps: obtaining a computing unit structure of each layer in each module serially serially connected in a depth neural network in a preset search space; the preset connection mode is adopted in each module to connect each computing unit to obtain the information flow in each module. According to the connection of the module and the computing unit in each module, the initial neural network is obtained. A sparse scaling operator is set up for the information flow in the initial neural network, in which the sparse scaling operator is used to scale the information flow. The preset training sample data is used to train the weights of the initial neural network and the sparse scaling operator of the information flow, and the intermediate neural network is obtained. The sparse scaling operator in the intermediate neural network is deleted to obtain the search result neural network in the search space. The invention can save the searching time of the network structure.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, in particular to a structure search method and device of a deep neural network. Background technique [0002] In recent years, deep neural networks have achieved great success in many fields, such as computer vision and natural language processing. Deep neural networks transform traditional hand-designed features into end-to-end learning through powerful representation capabilities. However, the structure of the current deep neural network is complicated, and there are many computing unit nodes such as convolution and pooling, so how to search for a model structure with a compact structure, fast running speed, and good effect in many computing unit nodes has become a problem. difficulty. [0003] At present, the prior art generally adopts to define a search space first, and then search for an optimal network structure in the search space. Generally, the heuristic method of netw...

AI Technical Summary

Problems solved by technology

[0004] Embodiments of the present invention provide a deep neural network structure search method and device to solve the problem that the time and calculation amount of network structure search in the prior art is extremely large. Problems with cumbersome and slow process

Images