The Bayesian Optimization of CNN Hyperparameters Based on Multi-Threaded Gaussian Process Acceleration
Hankun Yang, Yanmin Zhang, Chao Liu, Bowei Zhu · 2024
This study proposes a Bayesian optimization method that combines thread pooling with Gaussian process acceleration to optimize the hyperparameters of convolutional neural networks. In the experiment, this method showed significant advantages in hyperparameter tuning, as it can quickly find the optimal combination of hyperparameters and significantly reduce the validation set loss of the model. Compared to traditional Bayesian optimization strategies, the Bayesian optimization method combining Gaussian process acceleration and thread pooling reduces validation set loss by approximately 8.13%. By introducing a thread pool mechanism, computational efficiency has been improved, ensuring the stability of the model in a shorter period of time. This strategy has broad application prospects in the field of deep learning.