Date: December 17th, 12:00pm-1:00pm ET
Presenters: Weiguang Guan (SHARCNET, Alliance)
Modern AI models, especially deep neural networks, have achieved remarkable success across vision, language, and decision-making tasks — but their inner workings often remain opaque, earning them the label of “black boxes”. This lack of interpretability raises challenges in trust, accountability, and model debugging. In this talk, we explore Integrated Gradients, a principled method for attributing a model’s prediction to its input features. By integrating gradients of the model’s output with respect to its inputs along a path from a baseline to the actual input, this technique provides a mathematically grounded way to identify which features most influence the outcome. We will discuss the theoretical foundations of integrated gradients, their advantages over simpler attribution methods, and practical examples that illustrate how they reveal meaningful insights about model behavior.