PRUNING OF CONVOLUTIONAL NEURAL NETWORK ARCHITECTURE FOR CLASSIFICATION, REGRESSION, SEGMENTATION AND DETECTION TASKS

With the increasing size and complexity of neural network models the optimization of their performance has become an increasingly relevant challenge. Traditional pruning methods often face difficulties in maintaining accuracy or require significant computational resources. In this paper, a new pruning algorithm, U-Pruner, is introduced, which is designed to efficiently compress neural network models while preserving their high accuracy. In U-Pruner, neural network blocks are sequentially pruned, followed by fine-tuning of the best instances of the model. This approach allows insignificant parameters to be removed without losing key information, ensuring that an optimal balance between compression ratio and model accuracy is achieved. Experimental results have shown that U-Pruner outperforms traditional pruning methods on architectures such as MobileNetV2 and ResNet34. When models are compressed by 10 %, 50 %, and 80 %, high accuracy is maintained by U-Pruner, demonstrating the best results among its peers. Notably, when ResNet34 is compressed by 80 %, a decrease in model accuracy of only 3 % is observed, which significantly surpasses the results of other methods. Due to its versatility and efficiency, U-Pruner can be applied to a wide range of computer vision tasks, including classification, regression, segmentation, and detection. New opportunities for optimizing neural network models are opened by the proposed approach, making them more accessible for use in resource-constrained environments.