Hidden geometry of learning: Neural networks think alike