Neurosymbolic Learning

Neurosymbolic learning studies how to combine neural networks, which learn from data, with symbolic structures such as rules, logic, programs, and plans. Common approaches include differentiable logic and fuzzy constraints, probabilistic reasoning and answer set programming, and systems that learn rules or action models from examples. These methods are used in settings such as visual reasoning, robotics, process monitoring, scientific discovery, and classification with limited data or domain knowledge. A central challenge is making reasoning scalable and faithful: constraints can improve consistency and interpretability, but may hurt predictive performance or permit shortcuts that satisfy the rules without learning the intended concepts.