Diffusing Gaussian Mixtures for Generating Categorical Data

Learning a categorical distribution comes with its own set of challenges. A successful approach taken by state-of-the-art works is to cast the problem in continuous domain take advantage impressive performance generative models for data. Amongst them are recently emerging diffusion probabilistic models, which have observed generating high-quality samples. Recent advances focused on log likelihood improvements. In this work, we propose model data based focus sample generation, and sampled-based evaluation methods. The efficacy our method stems from performing while having parameterization informed structure nature target distribution. Our highlights capabilities limitations different data, includes experiments synthetic real-world protein datasets.