Pruning-Optimized Bayesian Neural Networks for Image Classification

A Pruned Bayesian Neural Network (PBNN) algorithm based on the collaborative optimization of variational inference and network pruning is proposed in this paper. It aims to address the issues of high computational complexity and model redundancy in traditional Bayesian Neural Networks (BNNs). The variational Bayesian framework is innovatively integrated with a parameter importance evaluation mechanism in the proposed method, and unstructured pruning is achieved through threshold determination. The uncertainty quantification advantage of BNNs is preserved while model efficiency is enhanced through this approach. In the MNIST digit classification benchmark, a classification accuracy of 97.49% is achieved by the proposed PBNN model. Additionally, the Brier Score of the pruned model is reduced by 46.25% compared to the unpruned model, demonstrating the superiority of the proposed algorithm in image recognition tasks. This study provides new insights for the efficient and practical application of Bayesian Neural Networks.