Less is More: The Influence of Pruning on the Explainability of CNNs

2023-02-17

arXiv.org

AI Summary

This study investigates the mechanistic impact of convolutional neural network (CNN) pruning on human interpretability—specifically, how model compression affects human understanding of model decisions. Method: We conduct a large-scale, controlled crowdsourcing experiment (37,500 annotation tasks) to systematically evaluate decision interpretability across 2×–32× compression ratios, integrating structured pruning, rigorously designed human-subject experiments, Amazon Mechanical Turk–based evaluation, and quantitative interpretability metrics. Contribution/Results: We provide the first empirical evidence of a non-monotonic relationship between pruning and human interpretability: moderate pruning (2×–4×) significantly enhances human comprehension of model reasoning, whereas aggressive pruning (32×) degrades explanation quality. A “sweet spot” emerges where classification accuracy and human-perceived interpretability are jointly optimized. These findings establish data-driven, empirically grounded pruning guidelines for explainable AI and delineate the fundamental trade-off between performance compression and cognitive understandability.

Abstract

Modern, state-of-the-art Convolutional Neural Networks (CNNs) in computer vision have millions of parameters. Thus, explaining the complex decisions of such networks to humans is challenging. A technical approach to reduce CNN complexity is network pruning, where less important parameters are deleted. The work presented in this paper investigates whether this technical complexity reduction also helps with perceived explainability. To do so, we conducted a pre-study and two human-grounded experiments, assessing the effects of different pruning ratios on CNN explainability. Overall, we evaluated four different compression rates (i.e., CPR 2, 4, 8, and 32) with 37 500 tasks on Mechanical Turk. Results indicate that lower compression rates have a positive influence on explainability, while higher compression rates show negative effects. Furthermore, we were able to identify sweet spots that increase both the perceived explainability and the model's performance.

Problem

• Network Pruning
• Convolutional Neural Networks
• Interpretability

Innovation

• Network Pruning
• Interpretable CNN
• Parameter Elimination