AFSD: Adaptive Feature Space Distillation for Distributed Deep Learning

Abstract

We propose a novel and adaptive feature space distillation method (AFSD) to reduce the communication overhead among distributed computers. The proposed method improves the Codistillation process by supporting longer update interval rates. AFSD performs knowledge distillates across the models infrequently and provides flexibility to the models in terms of exploring diverse variations in the training process. We perform knowledge distillation in terms of sharing the feature space instead of output only. Therefore, we also propose a new loss function for the Codistillation technique in AFSD. Using the feature space leads to more efficient knowledge transfer between models with a longer update interval rates. In our method, the models can achieve the same accuracy as Allreduce and Codistillation with fewer epochs.