In this talk, I will overview the state-of-the-art of logic-based machine learning for learning diferent classes of logic-based programs, (e.g., non-monotonic, non-deterministic and preference-based). I will then present how the advanced features of this family of systems can help the development of innovative neuro-symbolic AI solutions that combine statistical learning for fast “low-level” perception from unstructured data, with “high-level” symbolic learning of interpretable knowledge in a range of tasks. I will show that our neuro-symbolic solutions outperform diferentiable baseline systems in accuracy and data eficiency.