We propose a differentiable structure learning framework for general binary data that makes no parametric assumptions about the data‐generating process.