arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2409.04778cs.CLcs.LG
收藏

LoCa:面向知识蒸馏的 Logit 校准

LoCa: Logit Calibration for Knowledge Distillation

Tsinghua University(清华大学);The University of Hong Kong(香港大学)

Runming Yang, Taiqiang Wu, Yujiu Yang

AI总结:

针对知识蒸馏中教师 logits 与标签冲突导致的错误指导问题,本文提出无需额外参数的 LoCa,依据真实标签校准教师 logits,同时保留类别区分度等暗知识,并提升图像分类与文本生成基线性能。

摘要:

知识蒸馏(Knowledge Distillation,KD)旨在通过模仿教师模型来训练更优的学生模型,在模型压缩中发挥着重要作用。一种典型方法是对齐输出 logits。然而,我们发现了一个被称为错误指导(mis-instruction)的常见问题:当基于教师 logits 的预测与标签不一致时,学生会受到误导。同时,logits 中还存在其他有用的暗知识,例如类别区分度,这对蒸馏至关重要。本文提出了一种简单而有效的 Logit Calibration(LoCa)方法,根据真实标签校准来自教师模型的 logits。其核心洞见是在纠正预测(以解决错误指导问题)的同时保留有用的暗知识。所提出的 LoCa 不需要任何额外参数。在图像分类和文本生成任务上的实验结果表明,LoCa 能够有效提升基线方法的性能。

英文摘要:

Knowledge Distillation (KD), aiming to train a better student model by mimicking the teacher model, plays an important role in model compression. One typical way is to align the output logits. However, we find a common issue named mis-instruction, that the student would be misled when the predictions based on teacher logits do not follow the labels. Meanwhile, there is other useful dark knowledge in the logits such as the class discriminability, which is vital for distillation. In this paper, we propose a simple yet effective Logit Calibration (LoCa) method, which calibrates the logits from the teacher model based on the ground-truth labels. The key insight is to correct the prediction (to address the mis-instruction issue) and maintain useful dark knowledge simultaneously. Our proposed LoCa does not require any additional parameters. Empirical results on image classification and text generation tasks demonstrate that LoCa can effectively improve the performance of baselines.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑