Reducing the Teacher-Student Gap via Adaptive Temperatures

Knowledge distillation aims to obtain a small and effective deep model (student) by learning the output from a larger model (teacher). Previous studies found a severe degradation problem, that student performance would degrade unexpectedly when distilled from oversized teachers. It is well known that larger models tend to have sharper outputs. Based on this observation, we found that the sharpness gap between the teacher and student output may cause this degradation problem. To solve this problem, we first propose a metric to quantify the sharpness of the model output. Based on the second-order Taylor expansion of this metric, we propose Adaptive Temperature Knowledge Distillation (ATKD), which automatically changes the temperature of the teacher and the student, to reduce the sharpness gap. We conducted extensive experiments on CIFAR100 and ImageNet and achieved significant improvements. Specifically, ATKD trained the best ResNet18 model on ImageNet as we knew (73.0% accuracy).