On the relevance of logic for AI, and the promise of neuro-symbolic learning