Interpretable Learning Method Based on Causal Interactive Attention

The "black box" phenomenon inherent in deep neural networks poses significant challenges to the reliability and trustworthiness of their decision-making processes, particularly in high-stakes applications such as medical diagnosis and financial security. Current deep neural network models primarily use gradients or intermediate features to improve decision reliability, but most of the explanatory methods rooted in these features are post-hoc, with little direct influence on the classifiers’ decision outcomes. In contrast to human object recognition, where decisions and the underlying visual logic are clarified by both positive and negative causal evidence, deep neural networks often fail to achieve this level of nuanced understanding. To mitigate these limitations, we introduce a novel deep model based on Causal Interaction Attention (CIA), which aims to improve the reliability and interpretability of visual classification. Specifically, we first propose an interactive attention-driven decision mechanism that dynamically integrates visual features into the decision process, ensuring that each classifier output is accompanied by corresponding positive or negative visual interpretations. This mechanism not only explains "why this image belongs to a certain category", but also rationalizes "why this image does not belong to other categories", providing a comprehensive view of the model’s decision. In addition, to dynamically adapt to changes in the learning scenarios, we designed a loss function to adaptively control the strength of attention. It enhances the model to dynamically identify the most relevant visual cues, tailored to the granularity of the visual cues. The final experimental results show that the CIA can provide better visual interpretation while maintaining accuracy and reliability on small and medium-sized datasets.