BERT Learns to Teach: Knowledge Distillation with Meta Learning
This work addresses the challenge of efficient model compression for practitioners by making knowledge distillation more robust and applicable across different tasks and models, though it is incremental in nature.
The paper tackles the problem of improving knowledge distillation by enabling the teacher model to learn how to better transfer knowledge to the student, resulting in significant performance gains and reduced sensitivity to hyperparameters across various benchmarks.
We present Knowledge Distillation with Meta Learning (MetaDistil), a simple yet effective alternative to traditional knowledge distillation (KD) methods where the teacher model is fixed during training. We show the teacher network can learn to better transfer knowledge to the student network (i.e., learning to teach) with the feedback from the performance of the distilled student network in a meta learning framework. Moreover, we introduce a pilot update mechanism to improve the alignment between the inner-learner and meta-learner in meta learning algorithms that focus on an improved inner-learner. Experiments on various benchmarks show that MetaDistil can yield significant improvements compared with traditional KD algorithms and is less sensitive to the choice of different student capacity and hyperparameters, facilitating the use of KD on different tasks and models.