Neural network architecture search method, system and device and readable storage medium

Abstract

The invention discloses a neural network architecture search method, system and device and a readable storage medium, and the method comprises the steps: initializing the related parameters of a DARTS network, inputting an image training set into the initialized DARTS network, calculating a loss value according to a target function, calculating the network loss change through employing a second-order Taylor expansion according to gradient information, and obtaining the network loss change. The index saliency is calculated by using a grading index based on synaptic saliency, a connection sensitivity index is adopted to search a neural network architecture to indicate the importance of operation, microarchitecture structure search is defined as network pruning during initialization, operation saliency measurement is adopted in the initialization of the network pruning, and experimental results show that the operation significance of the neural network architecture is improved. The framework is a promising and reliable micro neural structure search solution, and good performance is achieved on different reference data sets and DARTS search spaces. The method is very efficient, and architecture search can be completed within several seconds.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence, and relates to a neural network architecture search method, system, equipment and readable storage medium. Background technique [0002] The success of deep learning in computer vision is largely due to the deep prior expertise of human experts, yet such manual design is costly and becomes increasingly difficult as networks grow larger and more complex. It is getting more and more difficult. Neural Network Architecture Search (NAS) automates the neural network design process and thus has received a lot of attention. However, this method requires very high computing power, and early NAS methods need to spend thousands of GPUs to find efficient network architectures. To improve efficiency, many recent studies have turned to reduce the cost of search, and one of the most popular paradigms is called Differentiable Neural Architecture Search (DARTS) framework. DARTS employs continuo...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to provide a neural network architecture search method, system, device and readable storage medium to solve the problem that the existing DARTS algorithm is too heavy in search cost and the search index cannot reflect the importance of operation

Images