Reproduce “Same Predictions, Different Reasons: The Effect of Quantization on Model Explanations” on OpenResearch

About this paper

Kazi Kamruzzaman Rabbi, Md. Zami Al Zunaed Farabe, M. Sohel Rahman

Researchers from BRAC University and Bangladesh University of Engineering and Technology reveal that while post-training quantization (PTQ) largely preserves deep learning model accuracy, it can significantly alter internal reasoning processes as shown by explainability methods. The study systematically evaluates various CNN architectures and precision levels, finding interpretability stability highly dependent on model architecture and the degree of quantization (INT8 vs. INT4).

Key insights

• Classification accuracy is not a reliable indicator of interpretability stability; models can maintain accuracy post-quantization while exhibiting significantly different reasoning patterns.
• Interpretability stability under PTQ is highly architecture-dependent: DenseNet161 demonstrated strong robustness, preserving explanations even at INT4, while EfficientNet-B0 was fragile, with severe degradation of input-level attributions.
• While lower precision (INT4) generally causes more pronounced interpretability degradation than INT8, the faithfulness (causal relevance) of explanations to the model's *own* predictions can remain preserved despite shifts in specific attribution patterns.

Summary by alphaXiv.