Pseudo Knowledge Distillation: Towards Learning Optimal Instance-specific Label Smoothing Regularization
Knowledge Distillation (KD) is an algorithm that transfers the knowledge of a trained, typically larger, neural network into another model under training. Although a complete understanding of KD is elusive, a growing body of work has shown that the success of both KD and label smoothing comes from a similar regularization effect of soft targets. In this work, we propose an instance-specific label smoothing technique, Pseudo-KD, which is efficiently learnt from the data. We devise a two-stage optimization problem that leads to a deterministic and interpretable solution for the optimal label smoothing. We show that Pseudo-KD can be equivalent to an efficient variant of self-distillation techniques, without the need to store the parameters or the output of a trained model. Finally, we conduct experiments on multiple image classification (CIFAR-10 and CIFAR-100) and natural language understanding datasets (the GLUE benchmark) across various neural network architectures and demonstrate that our method is competitive against strong baselines.
PDF Abstract
CIFAR-10
CIFAR-100
GLUE
QNLI
