How can rolling a ball downhill solve important multivariate problems? And what does it have to do with the core idea behind neural networks?