Error-correcting codes have been designed for decades using algebraic and combinatorial techniques. Recently, machine learning (deep learning) is used to construct error-correcting codes ``from scratch''. Are these codes similar to classical ones? We discuss some experiments exploring this question. The experiments use notions and techniques from combinatorics and theoretical computer science, such as influence, MILP, property testing and the Fourier representation of Boolean functions.
No prerequisites are assumed from coding theory or deep learning.
Joint work with N. Devroye, N. Mohammadi, A. Mulgund, H. Naik, R. Shekhar, Y. Wei and M. Zefran