Cross-Entropy is All You Need to Invert the Data Generating Process

This paper explores the effectiveness of supervised learning, demonstrating that models learn interpretable representations of variation through linear transformations, supported by theoretical insights and empirical studies.