Unification of Intelligence and Information: Beyond Blahut-Arimoto Using Machine Learning

The Blahut-Arimoto (BA) algorithm has been the cornerstone of computing channel capacity for discrete memoryless channels since 1972. In this work, we reveal a fundamental structural connection between the BA algorithm and the Rosenblatt perceptron learning algorithm. We show that BA can be interpreted as a specialized single-layer perceptron operating on channel transition matrices viewed as two-dimensional image data. Building on this insight, we propose a multilayer extension of the BA algorithm inspired by deep learning architectures, enabling the computation of cascaded channel capacities with explicit verification of the data processing inequality. Our framework opens new avenues for applying modern machine learning techniques to classical information-theoretic problems and suggests novel approaches to joint optimization of communication systems with multiple processing stages.