Interpretable Neural Networks: Bridging GAMs and Deep Learning

As machine learning models grow in complexity, interpretability often takes a backseat to predictive accuracy. While this tradeoff may be acceptable in some domains, it poses significant challenges in high-stakes fields such as healthcare, finance, and law, where model transparency is critical. Traditional methods, such as Generalized Additive Models (GAMs), offer high interpretability but suffer from limited flexibility and scalability. In contrast, deep learning models provide superior accuracy but often lack transparency. This thesis introduces a novel approach that fuses the interpretability of GAMs with the adaptability of neural networks. By leveraging neural network architectures, I propose an open-source framework that emulates the structural simplicity of GAMs while preserving the expressive power of modern deep learning methods. The method demonstrates competitive performance with state-of-the-art techniques in terms of accuracy while offering greater insight into model behavior. I provide empirical evaluations on benchmark datasets, highlighting the advantages and limitations of the framework, and propose directions for future research to refine this balance between transparency and predictive power.