Explainability of CNN Based Classification Models for Acoustic Signal

High accuracy alone isn't enough: The model achieved 94.8% accuracy, but the researchers found that understanding how the model made its decisions required additional explainability tools—demonstrating that performance metrics don't tell the full story about whether an AI system is trustworthy.

Different explanation methods reveal different insights: The study compared four different explainability techniques (LIME, SHAP, DeepLIFT, and Grad-CAM) and found they produced "different but complementary explanations," suggesting that relying on a single explanation method could give an incomplete or misleading picture of model behavior.

Combining multiple XAI approaches improves interpretability: When the researchers used multiple explainability techniques together, they gained "more complete and interpretable insights into the model's decision-making," indicating that responsible AI practice should involve triangulating across different explanation methods rather than depending on one.

XAI gaps exist even in specialized applications: The paper notes that while explainability tools have been applied across various acoustic domains, their use in bioacoustics (analyzing animal vocalizations) "remains relatively underexplored," highlighting that responsible AI practices may lag in niche but important application areas.

Explainability supports broader trust and interoperability: The authors argue that combining XAI techniques improves "trust and interoperability," suggesting that transparency tools are essential not just for understanding individual predictions but for building systems that different stakeholders can verify and work with.