Example-Based Explainable AI and its Application for Remote Sensing Image Classification

Training data quality directly impacts model reliability: The paper demonstrates that verifying what examples a model learned from is crucial for determining whether it will make trustworthy predictions. By showing similar training examples alongside predictions, users can confirm the model wasn't trained on fundamentally different data than what it's being asked to analyze.

Example-based explanations help non-experts assess AI decisions: The "What I Know" method makes AI more transparent by displaying actual training examples similar to new inputs, allowing stakeholders without technical expertise to visually verify whether a model's decision makes sense rather than relying solely on confidence scores or abstract explanations.

Context matters more than simple similarity: The paper emphasizes that selected examples should be similar to input data "in the context of the model task," not just statistically similar. This means explanations are most useful when they show relevant examples that actually relate to what the model is trying to accomplish.

Data-centric approaches complement model-focused AI development: Rather than only improving model architecture, this work highlights the importance of examining training datasets themselves. Organizations can use this method to identify gaps in their training data and determine whether additional data collection is needed before deploying models in real-world applications.

The method is broadly applicable across domains: While demonstrated with satellite imagery classification, the authors note this approach can extend to various machine learning tasks including regression, suggesting it could help verify AI reliability across different industries and use cases.