Bi-Tempered Logistic Loss for Training Neural Nets with Noisy Dataai.googleblog.com1 pointtheafh7 years ago