Spectral Sparsity: A Unifying Framework for Scalable Model Interpretability Using Codes

As machine learning models grow in complexity and are increasingly deployed in high-stakes applications, the need for tools to interpret their outputs is essential for ensuring their trustworthiness. Existing interpretability methods often fail to scale to modern models because of their high computational cost, or are too simplistic and fail to capture higher-order interactions. This tutorial presents a unifying framework that addresses these challenges, leveraging coding theory to exploit the natural spectral sparsity many models exhibit under the Fourier and Möbius transforms. Algorithms derived from algebraic and sparse-graph codes can be used to efficiently recover this sparse support, revealing important feature interactions across a variety of tasks and modalities. We showcase this framework's utility on applications ranging from uncovering complex linguistic patterns in large language models to identifying epistatic interactions in biological sequence models.