Quantization impacts deep learning model explanations, study finds

A new study published on arXiv investigates how post-training quantization (PTQ) affects the explainability of deep learning models. Researchers evaluated five common CNN architectures (VGG19, ResNet18, EfficientNet-B0, DenseNet161, and MobileNetV2) at INT8 and INT4 precision using a dual framework combining Grad-CAM and LIME. The findings indicate that quantization can alter a model's internal reasoning, with DenseNet161 showing better interpretability consistency than EfficientNet-B0, despite similar classification accuracy. This highlights the importance of considering architecture choice alongside quantization strategy for trustworthy deployment in applications requiring high interpretability. AI

IMPACT Quantization strategies can degrade model interpretability, necessitating careful architecture selection for trustworthy AI deployments.

RANK_REASON Research paper detailing experimental findings on model explainability. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• DenseNet161
• EfficientNet B0
• Grad-CAM++
• Kazi Kamruzzaman Rabbi
• MobileNetV2
• ResNet18
• VGG19

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →