On the nature of theories: a neurocomputational perspective