Automated Neural Network Architecture Design and Hyperparameter Adjustment Integrating Distribution-Based Optimization Strategy

Collaborative design and optimization of neural network structure and hyperparameter is the key to improving the efficiency of deep learning model. Conventional methods such as manual tuning and grid search often consume large computational resources and have limited efficiency when facing complex parameter combinations. In this paper, an automated scheme based on probability distribution modeling is developed to realize the synchronous optimization of network architecture design and hyperparameter tuning. The method innovatively transforms various parameters into dynamically adjusted probability distributions and drives the evolution of distribution morphology through a continuous iterative performance feedback mechanism. Based on the intelligent exploration strategy of unified probability space, this scheme can effectively overcome the dimensional disaster problem of traditional methods in high-dimensional configuration space, and can quickly locate the high-performance model configuration under limited computational resources. Comparative experiments on cifar 10, Fashion-MNIST, and IMDB show that the proposed scheme is significantly superior to manual design, grid search, and Bayesian optimization in classification accuracy, model stability, and computing power utilization. This achievement provides a new technical direction for the field of machine learning, and its basic mechanism can be extended to many fields such as image recognition and text processing, which has practical value in promoting the efficiency of intelligent algorithm development.