Investigating the Effect of Network Pruning on Performance and Interpretability

Technical Summary: Investigating the Effect of Network Pruning on Performance and Interpretability

Problem Statement
Deep Neural Networks (DNNs), such as GoogLeNet, are frequently over-parameterized, creating challenges for deployment in resource-constrained environments. While neural network pruning is a established technique for reducing model size, existing literature has primarily focused on performance metrics like accuracy and space complexity. There is a notable gap in understanding how different pruning strategies affect model interpretability. Furthermore, while methods exist to quantify interpretability, such as the Mechanistic Interpretability Score (MIS), their reliability and alignment with intuitive notions of interpretability (e.g., understanding the basis of correct decisions) in the context of pruned networks remain under-explored.

Methodology
The authors systematically investigated the impact of pruning on GoogLeNet using three distinct strategies:

1. Unstructured Pruning: Removes individual weights across the network based on magnitude, creating irregular sparsity.
2. Structured Pruning: Removes entire filters or channels, resulting in regular sparsity patterns amenable to hardware acceleration.
3. Connection Sparsity: A novel approach for this study focusing on pruning input channels rather than output channels. This aims to preserve structural integrity while reducing connections.

The study evaluated these strategies under two retraining regimes:

• One-Shot Pruning: A large portion of the network is pruned in a single step, followed by retraining.
• Iterative Pruning: The network is pruned in smaller increments with retraining after each step, allowing gradual adaptation.

Experiments were conducted on the ImageNet validation set. The authors tested various hyperparameters, identifying SGD with an ExponentialLR scheduler and 50 training epochs as optimal for retraining. To assess interpretability, the authors employed the Mechanistic Interpretability Score (MIS) developed by Zimmermann et al., which measures the perceptual similarity between explanations (images maximizing filter activations) and queries.

Key Results

• Performance:

  • Connection Sparsity outperformed both unstructured and structured pruning. Notably, pruning 80% of the model's parameters via Connection Sparsity and retraining resulted in a validation accuracy of 70.5%, surpassing the default GoogLeNet's performance.
  • Iterative pruning consistently yielded higher accuracy than one-shot pruning, though it incurred a higher computational cost (requiring approximately 50 epochs per pruning step).
  • Contrary to the assumption that pruning degrades performance, the authors observed that with sufficient retraining, pruned networks could approximate or even exceed the accuracy of the unpruned baseline. For instance, a model with 60% of parameters pruned achieved 72.286% accuracy, compared to 72.1% for a standard GoogLeNet trained for the same additional epochs.
  • Unstructured pruning without fine-tuning resulted in drastic accuracy drops, emphasizing the necessity of retraining.

• Interpretability:

  • Using the MIS metric, the study found no significant relationship between the pruning rate and interpretability (correlation coefficient of 0.092).
  • A critical finding was that networks with extremely low accuracy (e.g., 0.1%) could still achieve high MIS scores. This suggests that MIS measures the consistency of unit activations rather than the correctness of the decision-making process. A network might be "interpretable" by MIS standards (knowing which units activate for specific inputs) while making random or incorrect predictions.
  • The MIS scores for the softmax layer were consistently low across all models, contradicting the expectation that class units should be highly interpretable. Furthermore, there was no significant correlation between class-wise accuracy and MIS scores in the softmax layer for the tested pruned models.

Key Contributions

1. Evaluation of Connection Sparsity: The paper introduces and validates "Connection Sparsity" (pruning input channels) as a highly effective pruning strategy for GoogLeNet, demonstrating its ability to reduce model size by up to 80% while maintaining or improving accuracy.
2. Comparative Analysis of Retraining Strategies: The study provides empirical evidence comparing iterative versus one-shot pruning, confirming the performance benefits of iterative approaches while highlighting their computational inefficiency.
3. Critical Assessment of MIS: The authors challenge the utility of the MIS as a standalone metric for interpretability in pruned networks. They demonstrate that high MIS scores do not necessarily correlate with high accuracy or meaningful decision-making, indicating a misalignment between the metric and intuitive interpretability.

Significance and Claims
The authors claim that their work contributes to the development of more efficient deep learning models by identifying Connection Sparsity as a superior pruning method for GoogLeNet. However, the paper is modest regarding its interpretability findings. It does not claim to have solved the problem of interpretability but rather highlights the limitations of existing metrics like MIS. The study suggests that current measures may fail to capture the nuances of model interpretability post-pruning, particularly when the network's predictive quality is low. Consequently, the authors argue for the development of more robust metrics that align interpretability scores with the quality of decisions made by the network, a necessity for deploying neural networks in critical real-world applications.