How do neural networks learn relationships within data, and how does this differ from relational learning in traditional AI approaches?