Title: InfoAtlas: A Foundational Model for Zero-Shot Statistical Dependence Estimation

Abstract: Determining the statistical dependence between high-dimensional random variables constitutes a core challenge in both data science and machine learning. While neural mutual information (MI) estimators present a viable solution, they are often hindered by the need for computationally expensive iterative optimization for every distinct dataset, rendering them unsuitable for real-time deployment. To address this limitation, we introduce InfoAtlas, a foundation model-style architecture capable of inferring MI instantaneously via a single forward pass. Trained on extensive synthetic datasets characterized by diverse dependence patterns, InfoAtlas acquires the ability to recognize various dependency structures and output MI predictions directly. Extensive experimental results show that InfoAtlas rivals state-of-the-art neural estimators in precision while delivering a $100\times$ improvement in speed. Furthermore, the model offers the flexibility to manage different sample sizes and dimensions through a unified framework and demonstrates robust generalization to complex, real-world contexts. By reframing MI estimation as an inference problem, InfoAtlas lays the groundwork for instantaneous dependency analysis.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC