Deep Learning with Requirements in the Real World

Abstract

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.