Understanding and Improving Knowledge Distillation for Quantization-Aware Training of Large Transformer Encoders

TL;DR

This paper analyzes how knowledge distillation can be optimized for quantization-aware training of large Transformer models, proposing new methods that improve accuracy with ultra-low precision weights.

Contribution

It introduces attention-map and attention-output distillation losses and unifies them, enhancing QAT performance for large Transformers with sub-2-bit weights.

Findings

Proposed KD methods outperform previous approaches.

Achieved state-of-the-art accuracy in 2-bit quantized Transformers.

Improved attention recovery in quantized models.

Abstract

Knowledge distillation (KD) has been a ubiquitous method for model compression to strengthen the capability of a lightweight model with the transferred knowledge from the teacher. In particular, KD has been employed in quantization-aware training (QAT) of Transformer encoders like BERT to improve the accuracy of the student model with the reduced-precision weight parameters. However, little is understood about which of the various KD approaches best fits the QAT of Transformers. In this work, we provide an in-depth analysis of the mechanism of KD on attention recovery of quantized large Transformers. In particular, we reveal that the previously adopted MSE loss on the attention score is insufficient for recovering the self-attention information. Therefore, we propose two KD methods; attention-map and attention-output losses. Furthermore, we explore the unification of both losses to address task-dependent preference between attention-map and output losses. The experimental results on various Transformer encoder models demonstrate that the proposed KD methods achieve state-of-the-art accuracy for QAT with sub-2-bit weight quantization.