Interpretable Machine Learning

ML algorithms, and deep learning modelsModels more so, are notorious for their black-box nature providing little or no insights into the nature of the learned function. To address this challenge, increasing emphasis is being placed on developing interpretable ML modelsModels or enabling interpretability for the existing modelsModels. Interpretable machine learningMachine learning addresses the challenge of understanding complex black-box modelsModels, enabling transparency and insight into their decision-making processes. In this chapter, we explore interpretable machine learningMachine learning techniques, focusing on two prominent methods: SHapley Additive exPlanations (SHAP) and integrated gradients. The SHAP framework, based on cooperative game theory, is examined as a method to attribute feature contributions to modelModels outputs. SHAP values provide a mathematically grounded approach to understanding the significance and impact of individual features. The chapter also explores integrated gradients, which quantify feature importance by integrating gradients along a reference-to-input path. This technique offers insights into how changes in feature values affect modelModels predictions. We also discuss symbolic regression as a tool to abstract out symbolic laws from the data. Finally, we discuss a few additional interpretability algorithms to unpack the black-box ML modelsModels. Altogether, we discuss how interpretable algorithms can provide insights into the feature-to-label map learned by the DL modelsModels.