Explainable Artificial Intelligence for Bayesian Neural Networks: Towards trustworthy predictions of ocean dynamics

Bayesian Neural Networks can quantify uncertainty in AI predictions: Unlike standard neural networks that provide single-point predictions, this research demonstrates that Bayesian Neural Networks express predictions as probability distributions, allowing practitioners to calculate the statistical significance of outcomes—a critical capability for high-stakes applications like climate modeling.

Multiple explainability techniques provide complementary insights: The paper shows that using two different explanation methods (Layer-wise Relevance Propagation and SHAP values) together offers a more complete picture of how an AI model works than either method alone, since LRP reveals which network parameters matter most while SHAP shows how input changes affect outputs.

AI explanations can validate against established science: The researchers verified their explainability techniques by comparing them to known physical theory about ocean dynamics, demonstrating that trustworthy AI should produce explanations that align with domain expertise—and when they don't, this can identify gaps in current scientific understanding.

Transparency tools help determine when AI systems are reliable: By spatially applying explainability techniques across predictions, the paper shows practitioners can assess whether and where a neural network is trustworthy enough for real-world decision-making, rather than treating the model as a "black box."

Responsible AI requires addressing both uncertainty and interpretability: The research emphasizes that trustworthiness in high-stakes domains requires solving two problems simultaneously—knowing how confident the AI should be in its predictions AND understanding why it made those predictions.