Meta-learning Driven Automatic Hyperparameter Optimization for Neural Networks in Computer Vision

With the growing complexity of computer vision tasks and the explosive expansion of data scales, the optimization of neural network hyperparameters has become increasingly critical to model performance. Efficient hyperparameter optimization enables breakthroughs in tackling complex tasks and adapting to intricate scenarios. Traditional optimization methods rely heavily on manual experience and suffer from low efficiency. In contrast, meta-learning empowers hyperparameter optimization through its core logic of “learning to learn”: it extracts generalizable experience to build meta-cognition, thereby enhancing computational efficiency while achieving strong generalization capabilities. This paper first systematically introduces basic concepts such as neural networks, then elaborates on classic traditional optimization methods and meta-learning-based approaches, and then presents the experimental results of selected algorithms. Finally, it identifies unresolved issues and provides an outlook on future development trends.