The Neglected Baseline in Model Interpretation
Researchers have identified a critical oversight in current model interpretation techniques: the neglect of baselines. This paper argues that ignoring baselines leads to inaccurate or flawed interpretations of AI models. The authors propose a reformulated approach to model interpretation, unifying existing methods like gradient-based techniques and Taylor expansion, and explicitly defining baselines for each. They advocate for a new evaluation metric based on attribution error and introduce an improved interpretation method that achieves better results by incorporating a clear baseline. AI
IMPACT Introduces a more rigorous framework for understanding AI model behavior, potentially leading to more reliable AI systems.