Hierarchical Channel Clustering with Bayesian-Optimized Pruning for High-Fidelity CNN Compression

Parameter pruning in deep neural networks is a widely adopted technique for model compression. Conventional pruning methods primarily evaluate weight importance using predefined importance factors; however, they often result in significant accuracy degradation at high compression rates. An alternative strategy involves categorizing weights and selecting the most informative representatives within each category. Both approaches eliminate a substantial number of weights based on specific evaluation criteria, potentially overlooking their collective impact on model accuracy. To address this limitation, this study introduces a weight clustering and merging approach for compressing convolutional neural networks. A weight feature matrix is first constructed based on channel relationships, followed by hierarchical clustering using Ward's distance. The pruning ratio for each individual layer is adjusted through Bayesian optimization, followed by a reconstruction of weights guided by the geometric relationships among features. In comparison with recent leading approaches, our method attains superior performance in terms of accuracy. In certain parameter settings, it even outperforms the unpruned baseline model.