Core Concepts
DSEG-LIME introduces data-driven segmentation to enhance image explanation interpretability, outperforming conventional methods.
Abstract
The content discusses the introduction of DSEG-LIME, a framework that improves image explanation interpretability by incorporating data-driven segmentation. It addresses challenges in LIME's feature generation and hierarchical segmentation, showcasing superior performance in XAI metrics and user studies. The study evaluates DSEG-LIME against pre-trained models using ImageNet data, highlighting its effectiveness in enhancing model explanations.
Abstract:
Explainable AI (XAI) is crucial for understanding complex ML models.
LIME framework uses image segmentation for classification feature identification.
DSEG-LIME introduces data-driven segmentation for improved interpretability.
Outperforms in XAI metrics and aligns explanations with human concepts.
Introduction:
Integration of AI services requires trust and accuracy.
LIME aims to demystify AI decision-making through key feature identification.
Challenges arise from poor segmentation affecting explanation consistency.
Data-Driven Segmentation:
DSEG-LIME integrates SAM for improved feature quality.
Hierarchical segmentation allows adjustable granularity for tailored explanations.
Hierarchical Segmentation:
SAM enables fine and coarse segmentations for compositional object structures.
Users can specify granularity levels for detailed explanations.
Evaluation:
Quantitative assessment using established XAI metrics shows DSEG-LIME outperforms other methods.
User study confirms DSEG-LIME as the preferred explanation method.
Related Work:
Instability in LIME explanations addressed by SLIME, BayLIME, and GLIME.
Segmentation techniques significantly impact explanation quality and stability.
Conclusion:
DSEG-LIME enhances model interpretability through data-driven segmentation, promising more transparent AI insights. Further exploration with alternative foundational models is suggested for future work.
Stats
DSEG-LIMEは画像説明の解釈を向上させるためにデータ駆動セグメンテーションを導入します。