A New Link to an Old Model Could Crack the Mystery of Deep Learning