Research On Multi-objective Convolutional Neural Architecture Search Algorithm Based On Population Coevolution

The neural network simulates the function and connection structure of the neurons in the biological brain,so that the machine can automatically process feature engineering through data-driven.With the improvement of hardware computing power,machine intelligence technology represented by deep learning has been widely applied in industry and academia.As the performance of neural network highly depends on its internal complex topological structure.The design of network structure is developing towards complexity and large-scale.A large amount of relevant professional knowledge and tedious tuning tasks has hindered the practice and development of deep learning in different fields to a certain extent.Therefore,the automatic design of network architecture through computers instead of experts has become a research trend in the field of deep learning.However,although the related algorithm has achieved excellent results beyond the performance of manual networks,the search process requires a lot of time and calculation costs,which makes this technology unable to be widely promoted in various fields.Aiming at the challenges existing in neural architecture search,a multi-objective convolutional neural network architecture search method based on population coevolution is designed in this paper,which improves the search efficiency and reduces the computational cost while ensuring the accuracy.A more complete search space is defined in this algorithm,and a set of flexible gene encoding and decoding schemes are designed to improve the generation of offspring and the efficiency of individual evaluation in the iteration of the population.Regarding the high cost of super network training during the search process and the uneven training of sub-architectures with different scales,a population-based super-network sampling training strategy is proposed to effectively reduce the memory consumption and balance the convergence of sub-architectures.At the same time,batch architecture training strategy can effectively avoid coupling interference between architectures.In the population evaluation process,a strategy of simultaneous optimization of multiple conflicting objectives is adopted to instead of the single objective of best accuracy to evaluate the subarchitecture,high-performance architectures with different complexity can be designed for different deployment scenarios,while weight sharing mechanisms are used instead The weight training process avoids the time spent on heavy training and effectively improves the evaluation efficiency of environment selection.Through a multi-dimensional comprehensive comparison with excellent manual design architectures and peer algorithms on the image classification test benchmark,the accuracy performance exceeds most existing manual architectures.At the same time,in the comparison of peer algorithms,the search time is guaranteed under the premise of ensuring accuracy.Significantly reduced and less computational overhead.On the other hand,the proposed algorithm is better at discovering high-precision small-scale architectures in the search process,and the performance of the algorithm is verified.