https://baeseongsu.github.io/posts/knowledge-distillation/knowledge-distillation-pytorchhaitongli • Updated Dec 7, 2023performance 관점에서는 많은 모델을 사용하는 것이 좋음 → deploy 관점에서는 cost high.→ KD 제안. ExperimentMNIST에서 label 3 없애고 soft label만 학습시킴에도 86%얻음.