Deep learning models have repeatedly shown their strengths in various application domains. However, their predictions often struggle to meet background knowledge requirements, which is a crucial condition for safety-critical systems. My research focuses on integrating requirements into neural networks to guide the learning process and ultimately produce outputs that ensure the requirements' satisfaction. Here, I will discuss my proposed methods in the context of two real-world applications: tabular data generation and autonomous driving.