دانلود کتاب جستجوی معماری عصبی عمیق تکاملی: مبانی، روش‌ها و پیشرفت‌های اخیر

The use of Evolutionary Computation (EC) methods to create optimal or nearly optimal Deep Neural Network (DNN)architectures is referred to as evolutionary deep neural architecture design. The design process of architectures is often formalized as an optimization problem, where EC algorithms are correctly created to tackle the optimization problem. DNNs have had remarkable success in many complicated practical applications in recent years. It is well known that the performance of a DNN is only promising when the architecture is appropriate. The architecture, on the other hand, is typically created by hand, needing a high level of skill that is in scarce supply in practice. A promising deep architecture is difficult to create in practice without this kind of skill, which is frequently DNN expertise and domain understanding of the problem to be solved. Reinforcement learning techniques, gradient-based optimization algorithms, and EC methods are the three commonmethods utilized to construct the architectures of DNNs in the literature. This book mainly focuses on the EC methods for deep neural architecture design. In this book, we will first introduce the fundamentals of commonly used EC methods, including Genetic Algorithm (GA) [1], ParticleSwarmOptimization (PSO) [2], Differential Evolution (DE) [3], and Genetic Programming (GP) [4]. Following that, we will go through two different forms of evolutionary deep neural architecture design algorithms. They are the design algorithms for unsupervised DNNs, and the design algorithms for supervised DNNs. In addition, we will also discuss some recent efforts to speed up the execution of such algorithms. These algorithms are primarily based on the authors recent work, which has been published in journals and international conferences devoted to EC and neural networks. We think that by presenting them all together in this book, readers will be able to absorb related information more quickly and conveniently.