Multi-task neural network architecture searching method based on evolutionary computation

Abstract

The invention discloses a multi-task neural network architecture searching method based on evolutionary computation, which comprises the following steps: firstly, initializing a population; evaluating the multi-task generalization abilities of individuals in the population; then randomly obtaining two chromosomes through a binary tournament selection algorithm; comparing the multi-task generalization performance of the two chromosomes; selecting the chromosome with better performance as a parent; then carrying out crossover and mutation operations on two parents to generating two children; evaluating the multi-task generalization performance of the children; then combining the children and the parents; carrying out environment selection according to an evaluation result; generating a new population; carrying out a new round of evolution until a predetermined termination condition is reached; and outputting the individual with the best multi-task generalization ability. According to the method, a genetic algorithm is used for optimizing the multi-task network model system structure, the neural network model suitable for multi-task learning can be automatically searched out without manual participation, and the cross-task information fusion capability of the multi-task network is improved.

Description

technical field [0001] The invention relates to the field of neural networks, in particular to a search method for a multi-task neural network architecture based on evolutionary computation. Background technique [0002] At present, related technologies for single machine vision tasks, such as image classification and target recognition, whether it is the traditional non-neural network method or the deep learning method based on the convolutional neural network, have developed relatively mature. However, in real-world scenarios, it is often necessary to process multiple tasks at the same time. For example, unmanned driving requires road target recognition and target depth estimation to ensure the safe operation of unmanned vehicles. Multi-task learning aims to make full use of the information between multiple associated tasks to improve the generalization ability of the model and the performance of a single task. [0003] Multi-task learning techniques have continuously ev...

AI Technical Summary

Problems solved by technology

This method may lead to the fusion of some useless information, which will affect the performance of the model

Images