Self-adaptive Neuro Evolution Method For Deep Neural Networks In Computer Vision

The theoretical and technical advancements in the field of computer vision have found wide applications in academic research,industrial production,and everyday life.Recently,propelled by rapid developments in artificial intelligence,Deep Neural Networks(DNNs)have become a key technology in computer vision.As computer vision applications expand,efficiently constructing various types of DNNs has emerged as a forefront research topic.This has led to the development of neural architecture search(NAS)methods that can automatically construct DNN architectures.Existing methods,however,mainly focus on designing single-type DNN architectures and lack universality in terms of balancing DNN performance and scale.To address these issues,this study proposes a NAS method for automatically constructing various types of DNN architectures in the field of computer vision,named Self-Adaptive Neuro Evolution(SANE)due to its self-adaptive search space and evolutionary algorithm-based search strategy.The main research contents of this study are:1.Developing a three-tier evolution search space.This study analyzes the commonalities and differences among various DNN types used in computer vision tasks,integrating the high freedom of the global search space with the efficiency of the cell-based search space.The evolution search space is divided into three levels: micro-structural modules,meso-functional units,and macro-task regions.Depending on the design needs of different types of DNNs in computer vision tasks,this three-tier evolution search space can adaptively adjust its settings at each level to instantiate the evolution search space for the corresponding DNN type.Experimental results show that this search space can adaptively instantiate search spaces for various types of DNNs,including Convolutional Neural Networks,Generative Adversarial Networks,and You Only Look Once.2.Proposing a constructive evolution search strategy.This search strategy initiates with a population of identical,minimal DNN architectures and gradually increases the DNN scale through multi-granularity evolution operations at different levels to build minimal DNN architectures under required performance criteria.To address the issue of insufficient diversity in the initial evolution population,this strategy uses clustering algorithms to divide the evolution population into multiple distinct species based on genotypic distances,maintaining diversity through interspecies isolation and intraspecies competition.Experimental results indicate that this strategy can optimize the performance and scale of DNN architectures.3.Introducing a set of evolution search optimization scheme.The scheme includes a evolution hyperparameter dynamic adjustment strategy for evolutionary hyperparameters,which adjusts settings based on the state of the evolutionary search,ensuring process stability.An ageperformance hybrid selection operation addresses potential misguidance from imprecise performance evaluations,balancing exploration and exploitation in the evolution search.A approaching convergence performance evaluation mechanism optimizes the evaluation of DNN parameters toward performance convergence,significantly reducing computational resources and time costs.Experimental results demonstrate that this optimization scheme effectively improves the stability and efficiency of the evolution search.