TbExplain: A Text-based Explanation Method for Scene Classification Models with the Statistical Prediction Correction

Text explanations are more accessible than visual heatmaps: The paper identifies a key limitation of traditional heatmap-based explanations—non-expert users often struggle to understand why certain pixels are highlighted. TbExplain addresses this by providing text-based explanations that describe which objects and regions influenced the model's decision, making AI reasoning more understandable to the general public.

AI systems can be made more reliable through statistical correction: TbExplain includes a novel feature that detects when the model's initial prediction is unreliable and corrects it based on the statistical likelihood of objects appearing in the scene. This demonstrates that explainability methods can simultaneously improve both transparency and accuracy—in this case, showing improved classification accuracy over standard ResNet models.

Combining multiple AI techniques improves explanation quality: The framework integrates explainability techniques with object detection to provide more complete explanations. Rather than just highlighting relevant pixels, it identifies and names specific objects that influenced the decision, reducing the ambiguity that heatmaps can create about which regions actually matter.

Trustworthiness of AI explanations can be validated: The researchers conducted both qualitative and quantitative experiments to assess whether their text-based explanations were actually reliable and trustworthy. This methodological approach—testing whether explanations are genuinely faithful to the model's reasoning—is important for ensuring that explanations don't mislead users into false confidence in AI systems.