Self-Distillation as Instance-Specific Label Smoothing

           Zhilu Zhang                Mert R. Sabuncu
         Cornell University              Cornell Univerisity
         zz452@cornell.edu             msabuncu@cornell.edu

                       Abstract
     It has been recently demonstrated that multi-generational self-distillation can im-
     prove generalization [11]. Despite this intriguing observation, reasons for the
     enhancement remain poorly understood. In this paper, we f ...