Although deep neural networks are becoming increasingly powerful, they do not typically benefit from rules, which encode prior knowledge (e.g. heuristics, equations, constraints) efficiently and effectively and can be applied to improve accuracy and reliability. To enable data and rule-driven learning, Seo et al. propose DeepCTRL. DeepCTRL allows users to control the impact of rules on predictions (without retraining) by employing separate encoders for data and rules and combining the outputs stochastically. The authors use a perturbation-based method to transform non-differentiable rules into differentiable objectives.