Research on Group-level Pruning Algorithm Based on Progressive Sparsity Regularization

With the rapid development of autonomous driving systems, deep learning models play a crucial role in their implementation. However, these models typically have high computational complexity and storage demands, posing challenges to the real-time processing capabilities of embedded devices. To address this issue, this paper proposes a group-level pruning algorithm based on progressive sparsity regularization. The algorithm groups model parameters and progressively increases the sparsity regularization strength, effectively reducing computational load and storage requirements while maintaining model performance. Additionally, the algorithm introduces group-level weight importance calculation based on the L2-norm, combined with weight gradient direction information to achieve a more refined pruning strategy. Experimental results demonstrate that the proposed algorithm achieves a 2.58× speedup on the ResNet-56 model on the CIFAR dataset, significantly compressing the model's parameter size while maintaining accuracy. This research provides an effective solution for real-time applications such as autonomous driving, where efficient model deployment is essential.