Direct Optimization through $\arg \max$ for Discrete Variational Auto-Encoder

Models with discrete latent variables drive extensive research in machine learning applications, including language classification and generation [42, 11, 34], molecular synthesis [19], or game solving[25].