Abstract
Driven by the teacher-student paradigm, knowledge distillation is one of the de facto ways for language model compression. Recent studies have uncovered that conventional distillation is less effective when facing a large capacity gap between the teacher and the student, and introduced teacher assistant-based distillation to bridge the gap. As a connection, the scale and the performance of the teacher assistant is crucial for transferring the knowledge from the teacher to the student. However, existing teacher assistant-based methods manually select the scale of the teacher assistant, which fails to identify the teacher assistant with the optimal scale-performance tradeoff. To this end, we propose an Automatic Distillation Schedule (AutoDisc) for large language model compression. In particular, AutoDisc first specifies a set of teacher assistant candidates at different scales with gridding and pruning, and then optimizes all candidates in an once-for-all optimization with two approximations. The best teacher assistant scale is automatically selected according to the scale-performance tradeoff. AutoDisc is evaluated with an extensive set of experiments on a language understanding benchmark GLUE. Experimental results demonstrate the improved performance and applicability of our AutoDisc. We further apply AutoDisc on a language model with over one billion parameters and show the scalability of AutoDisc.
Abstract (translated)
URL
https://arxiv.org/abs/2205.14570
