Exploring Relevance-Based Pruning Strategies in VGG Models: A Comparative Study

This thesis investigates the effectiveness of different pruning strategies in VGG16 models. The study begins with the evaluation of the CIFAR-10 dataset using a VGG16 model, followed by the calculation of the model’s relevance using a method from the TorchLRP repository. Two pruning strategies were then implemented based on the calculated relevance data. The first strategy involved pruning the model based on the absolute mean value of the relevance of the channels. The second strategy used the absolute mean of the model parameters for pruning. Both methods were applied at varying pruning percentages, and their performances were compared. However, significant performance differences between the two methods were not observed until a combined approach was implemented. This combined method involved calculating the absolute mean value of both the relevance data and the parameters of the channels, sorting them, and alternating their order to form a new pruning list. The results showed that the combined pruning strategy allowed the model to maintain high accuracy and low loss, even at high pruning percentages, outper??forming the other two methods. This study provides valuable insights into the potential of relevance-based pruning strategies in enhancing the performance of VGG models