🤖 AI Summary
This study systematically investigates the dual impact of pruning on GoogLeNet’s performance and interpretability. We apply unstructured and structured pruning, input-connection sparsification, and both iterative and one-shot retraining strategies, quantifying interpretability via the Mechanistic Interpretability Score (MIS). Our key findings are: (1) With sufficient retraining, pruned models achieve ImageNet accuracy comparable to—or even exceeding—that of the original model; (2) MIS exhibits no significant correlation with either pruning ratio or classification accuracy, indicating that low-accuracy models can attain high MIS values. These results constitute the first empirical evidence that MIS fails to reliably reflect true mechanistic interpretability, thereby challenging prevailing score-based evaluation paradigms in interpretability research. The work establishes a new conceptual foundation for jointly optimizing model compression and interpretability, highlighting the need for more principled, behaviorally grounded metrics.
📝 Abstract
Deep Neural Networks (DNNs) are often over-parameterized for their tasks and can be compressed quite drastically by removing weights, a process called pruning. We investigate the impact of different pruning techniques on the classification performance and interpretability of GoogLeNet. We systematically apply unstructured and structured pruning, as well as connection sparsity (pruning of input weights) methods to the network and analyze the outcomes regarding the network's performance on the validation set of ImageNet. We also compare different retraining strategies, such as iterative pruning and one-shot pruning. We find that with sufficient retraining epochs, the performance of the networks can approximate the performance of the default GoogLeNet - and even surpass it in some cases. To assess interpretability, we employ the Mechanistic Interpretability Score (MIS) developed by Zimmermann et al. . Our experiments reveal that there is no significant relationship between interpretability and pruning rate when using MIS as a measure. Additionally, we observe that networks with extremely low accuracy can still achieve high MIS scores, suggesting that the MIS may not always align with intuitive notions of interpretability, such as understanding the basis of correct decisions.